benchmark_cases_metadata.csv CHANGED
@@ -1,19 +1,19 @@
1
- case_id,task_description,max_seq_len,min_seq_len,training_args.atol,training_args.lr,training_args.use_single_loss,training_args.iit_weight,training_args.behavior_weight,training_args.strict_weight,training_args.epochs,training_args.act_fn,training_args.clip_grad_norm,training_args.lr_scheduler,transformer_cfg.n_layers,transformer_cfg.d_model,transformer_cfg.n_ctx,transformer_cfg.d_head,transformer_cfg.model_name,transformer_cfg.n_heads,transformer_cfg.d_mlp,transformer_cfg.act_fn,transformer_cfg.d_vocab,transformer_cfg.eps,transformer_cfg.use_attn_result,transformer_cfg.use_attn_scale,transformer_cfg.use_split_qkv_input,transformer_cfg.use_hook_mlp_in,transformer_cfg.use_attn_in,transformer_cfg.use_local_attn,transformer_cfg.original_architecture,transformer_cfg.from_checkpoint,transformer_cfg.checkpoint_index,transformer_cfg.checkpoint_label_type,transformer_cfg.checkpoint_value,transformer_cfg.tokenizer_name,transformer_cfg.window_size,transformer_cfg.attn_types,transformer_cfg.init_mode,transformer_cfg.normalization_type,transformer_cfg.device,transformer_cfg.n_devices,transformer_cfg.attention_dir,transformer_cfg.attn_only,transformer_cfg.seed,transformer_cfg.initializer_range,transformer_cfg.init_weights,transformer_cfg.scale_attn_by_inverse_layer_idx,transformer_cfg.positional_embedding_type,transformer_cfg.final_rms,transformer_cfg.d_vocab_out,transformer_cfg.parallel_attn_mlp,transformer_cfg.rotary_dim,transformer_cfg.n_params,transformer_cfg.use_hook_tokens,transformer_cfg.gated_mlp,transformer_cfg.default_prepend_bos,transformer_cfg.dtype,transformer_cfg.tokenizer_prepends_bos,transformer_cfg.n_key_value_heads,transformer_cfg.post_embedding_ln,transformer_cfg.rotary_base,transformer_cfg.trust_remote_code,transformer_cfg.rotary_adjacent_pairs
2
- 11,Counts the number of words in a sequence based on their length.,10.0,4.0,0.05,0.01,False,1.0,1.0,0.4,500.0,gelu,1.0,,2.0,12.0,10.0,3.0,custom,4.0,48.0,gelu,10.0,1e-05,True,True,True,True,False,False,,False,,,,,,,gpt2,,cpu,1.0,causal,False,0.0,0.1460593486680443,True,False,standard,False,5.0,False,,3456.0,False,False,True,torch.float32,,,False,10000.0,False,False
3
- 13,"Analyzes the trend (increasing, decreasing, constant) of numeric tokens.",10.0,4.0,0.05,0.01,False,1.0,1.0,0.4,500.0,gelu,1.0,,2.0,20.0,10.0,5.0,custom,4.0,80.0,gelu,5.0,1e-05,True,True,True,True,False,False,,False,,,,,,,gpt2,,cpu,1.0,bidirectional,False,0.0,0.1460593486680443,True,False,standard,False,3.0,False,,9600.0,False,False,True,torch.float32,,,False,10000.0,False,False
4
- 18,"Classify each token based on its frequency as 'rare', 'common', or 'frequent'.",10.0,4.0,0.05,0.001,False,1.0,1.0,0.4,2000.0,gelu,0.1,,2.0,12.0,10.0,3.0,custom,4.0,48.0,gelu,7.0,1e-05,True,True,True,True,False,False,,False,,,,,,,gpt2,,cpu,1.0,bidirectional,False,0.0,0.12344267996967354,True,False,standard,False,3.0,False,,3456.0,False,False,True,torch.float32,,,False,10000.0,False,False
5
- 19,Removes consecutive duplicate tokens from a sequence.,15.0,4.0,0.05,0.001,False,1.0,1.0,0.4,2000.0,gelu,0.1,,2.0,32.0,15.0,8.0,custom,4.0,128.0,gelu,5.0,1e-05,True,True,True,True,False,False,,False,,,,,,,gpt2,,cpu,1.0,causal,False,0.0,0.15689290811054724,True,False,standard,False,3.0,False,,24576.0,False,False,True,torch.float32,,,False,10000.0,False,False
6
- 20,Detect spam messages based on appearance of spam keywords.,10.0,4.0,0.05,0.001,False,1.0,1.0,1.0,2000.0,gelu,0.1,,2.0,4.0,10.0,1.0,custom,4.0,16.0,gelu,14.0,1e-05,True,True,True,True,False,False,,False,,,,,,,gpt2,,cuda,1.0,causal,False,0.0,0.16,True,False,standard,False,2.0,False,,384.0,False,False,True,torch.float32,,,False,10000.0,False,False
7
- 21,Extract unique tokens from a string,10.0,4.0,0.05,0.01,False,1.0,1.0,0.4,500.0,gelu,1.0,,2.0,20.0,10.0,5.0,custom,4.0,80.0,gelu,5.0,1e-05,True,True,True,True,False,False,,False,,,,,,,gpt2,,cpu,1.0,causal,False,0.0,0.1885618083164127,True,False,standard,False,3.0,False,,9600.0,False,False,True,torch.float32,,,False,10000.0,False,False
8
- 24,Identifies the first occurrence of each token in a sequence.,10.0,4.0,0.05,0.01,False,1.0,1.0,0.4,500.0,gelu,1.0,,2.0,20.0,10.0,5.0,custom,4.0,80.0,gelu,5.0,1e-05,True,True,True,True,False,False,,False,,,,,,,gpt2,,cpu,1.0,causal,False,0.0,0.1885618083164127,True,False,standard,False,3.0,False,,9600.0,False,False,True,torch.float32,,,False,10000.0,False,False
9
- 3,Returns the fraction of 'x' in the input up to the i-th position for all i.,5.0,4.0,0.05,0.001,False,1.0,1.0,10.0,2000.0,gelu,0.1,,2.0,12.0,5.0,3.0,custom,4.0,48.0,gelu,6.0,1e-05,True,True,True,True,False,False,,False,,,,,,,gpt2,,cpu,1.0,causal,False,0.0,0.22188007849009167,True,False,standard,False,1.0,False,,3456.0,False,False,True,torch.float32,,,False,10000.0,False,False
10
- 33,Checks if each token's length is odd or even.,10.0,4.0,0.05,0.001,False,1.0,1.0,0.4,2000.0,gelu,0.1,,2.0,4.0,10.0,1.0,custom,4.0,16.0,gelu,10.0,1e-05,True,True,True,True,False,False,,False,,,,,,,gpt2,,cpu,1.0,causal,False,0.0,0.17457431218879393,True,False,standard,False,2.0,False,,384.0,False,False,True,torch.float32,,,False,10000.0,False,False
11
- 34,Calculate the ratio of vowels to consonants in each word.,10.0,4.0,0.05,0.001,False,1.0,1.0,0.4,2000.0,gelu,0.1,,2.0,4.0,10.0,1.0,custom,4.0,16.0,gelu,10.0,1e-05,True,True,True,True,False,False,,False,,,,,,,gpt2,,cpu,1.0,causal,False,0.0,0.16329931618554522,True,False,standard,False,5.0,False,,384.0,False,False,True,torch.float32,,,False,10000.0,False,False
12
- 35,Alternates capitalization of each character in words.,10.0,4.0,0.05,0.001,False,1.0,1.0,0.4,2000.0,gelu,0.1,,2.0,4.0,10.0,1.0,custom,4.0,16.0,gelu,10.0,1e-05,True,True,True,True,False,False,,False,,,,,,,gpt2,,cpu,1.0,causal,False,0.0,0.1539600717839002,True,False,standard,False,8.0,False,,384.0,False,False,True,torch.float32,,,False,10000.0,False,False
13
- 36,"Classifies each token as 'positive', 'negative', or 'neutral' based on emojis.",10.0,4.0,0.05,0.001,False,1.0,1.0,10.0,2000.0,gelu,0.1,,2.0,4.0,10.0,1.0,custom,4.0,16.0,gelu,5.0,1e-05,True,True,True,True,False,False,,False,,,,,,,gpt2,,cuda,1.0,causal,False,0.0,0.19402850002906638,True,False,standard,False,3.0,False,,384.0,False,False,True,torch.float32,,,False,10000.0,False,False
14
- 37,Reverses each word in the sequence except for specified exclusions.,10.0,4.0,0.05,0.001,False,1.0,1.0,0.4,2000.0,gelu,0.1,,2.0,4.0,10.0,1.0,custom,4.0,16.0,gelu,10.0,1e-05,True,True,True,True,False,False,,False,,,,,,,gpt2,,cpu,1.0,causal,False,0.0,0.1539600717839002,True,False,standard,False,8.0,False,,384.0,False,False,True,torch.float32,,,False,10000.0,False,False
15
- 38,Checks if tokens alternate between two types.,10.0,4.0,0.05,0.001,False,1.0,1.0,0.4,2000.0,gelu,0.1,,2.0,20.0,10.0,5.0,custom,4.0,80.0,gelu,5.0,1e-05,True,True,True,True,False,False,,False,,,,,,,gpt2,,cpu,1.0,causal,False,0.0,0.1539600717839002,True,False,standard,False,2.0,False,,9600.0,False,False,True,torch.float32,,,False,10000.0,False,False
16
- 4,Return fraction of previous open tokens minus the fraction of close tokens.,10.0,4.0,0.05,0.001,False,1.0,1.0,0.4,2000.0,gelu,0.1,,2.0,20.0,10.0,5.0,custom,4.0,80.0,gelu,7.0,1e-05,True,True,True,True,False,False,,False,,,,,,,gpt2,,cpu,1.0,causal,False,0.0,0.17056057308448835,True,False,standard,False,1.0,False,,9600.0,False,False,True,torch.float32,,,False,10000.0,False,False
17
- 8,Fills gaps between tokens with a specified filler.,10.0,4.0,0.05,0.01,False,1.0,1.0,0.4,500.0,gelu,1.0,,2.0,20.0,10.0,5.0,custom,4.0,80.0,gelu,10.0,1e-05,True,True,True,True,False,False,,False,,,,,,,gpt2,,cpu,1.0,causal,False,0.0,0.13333333333333333,True,False,standard,False,8.0,False,,9600.0,False,False,True,torch.float32,,,False,10000.0,False,False
18
- ioi,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,
19
- ioi_next_token,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,
 
1
+ case_id,url,task_description,max_seq_len,min_seq_len,training_args.atol,training_args.lr,training_args.use_single_loss,training_args.iit_weight,training_args.behavior_weight,training_args.strict_weight,training_args.epochs,training_args.act_fn,training_args.clip_grad_norm,training_args.lr_scheduler,transformer_cfg.n_layers,transformer_cfg.d_model,transformer_cfg.n_ctx,transformer_cfg.d_head,transformer_cfg.model_name,transformer_cfg.n_heads,transformer_cfg.d_mlp,transformer_cfg.act_fn,transformer_cfg.d_vocab,transformer_cfg.eps,transformer_cfg.use_attn_result,transformer_cfg.use_attn_scale,transformer_cfg.use_split_qkv_input,transformer_cfg.use_hook_mlp_in,transformer_cfg.use_attn_in,transformer_cfg.use_local_attn,transformer_cfg.original_architecture,transformer_cfg.from_checkpoint,transformer_cfg.checkpoint_index,transformer_cfg.checkpoint_label_type,transformer_cfg.checkpoint_value,transformer_cfg.tokenizer_name,transformer_cfg.window_size,transformer_cfg.attn_types,transformer_cfg.init_mode,transformer_cfg.normalization_type,transformer_cfg.device,transformer_cfg.n_devices,transformer_cfg.attention_dir,transformer_cfg.attn_only,transformer_cfg.seed,transformer_cfg.initializer_range,transformer_cfg.init_weights,transformer_cfg.scale_attn_by_inverse_layer_idx,transformer_cfg.positional_embedding_type,transformer_cfg.final_rms,transformer_cfg.d_vocab_out,transformer_cfg.parallel_attn_mlp,transformer_cfg.rotary_dim,transformer_cfg.n_params,transformer_cfg.use_hook_tokens,transformer_cfg.gated_mlp,transformer_cfg.default_prepend_bos,transformer_cfg.dtype,transformer_cfg.tokenizer_prepends_bos,transformer_cfg.n_key_value_heads,transformer_cfg.post_embedding_ln,transformer_cfg.rotary_base,transformer_cfg.trust_remote_code,transformer_cfg.rotary_adjacent_pairs
2
+ 11,https://huggingface.co/cybershiptrooper/InterpBench/tree/main/11,Counts the number of words in a sequence based on their length.,10,4,0.05,0.01,False,1.0,1.0,0.4,500.0,gelu,1.0,,2.0,12.0,10.0,3.0,custom,4.0,48.0,gelu,10.0,1e-05,True,True,True,True,False,False,,False,,,,,,,gpt2,,cpu,1.0,causal,False,0.0,0.1460593486680443,True,False,standard,False,5.0,False,,3456.0,False,False,True,torch.float32,,,False,10000.0,False,False
3
+ 13,https://huggingface.co/cybershiptrooper/InterpBench/tree/main/13,"Analyzes the trend (increasing, decreasing, constant) of numeric tokens.",10,4,0.05,0.01,False,1.0,1.0,0.4,500.0,gelu,1.0,,2.0,20.0,10.0,5.0,custom,4.0,80.0,gelu,5.0,1e-05,True,True,True,True,False,False,,False,,,,,,,gpt2,,cpu,1.0,bidirectional,False,0.0,0.1460593486680443,True,False,standard,False,3.0,False,,9600.0,False,False,True,torch.float32,,,False,10000.0,False,False
4
+ 18,https://huggingface.co/cybershiptrooper/InterpBench/tree/main/18,"Classify each token based on its frequency as 'rare', 'common', or 'frequent'.",10,4,0.05,0.001,False,1.0,1.0,0.4,2000.0,gelu,0.1,,2.0,12.0,10.0,3.0,custom,4.0,48.0,gelu,7.0,1e-05,True,True,True,True,False,False,,False,,,,,,,gpt2,,cpu,1.0,bidirectional,False,0.0,0.12344267996967354,True,False,standard,False,3.0,False,,3456.0,False,False,True,torch.float32,,,False,10000.0,False,False
5
+ 19,https://huggingface.co/cybershiptrooper/InterpBench/tree/main/19,Removes consecutive duplicate tokens from a sequence.,15,4,0.05,0.001,False,1.0,1.0,0.4,2000.0,gelu,0.1,,2.0,32.0,15.0,8.0,custom,4.0,128.0,gelu,5.0,1e-05,True,True,True,True,False,False,,False,,,,,,,gpt2,,cpu,1.0,causal,False,0.0,0.15689290811054724,True,False,standard,False,3.0,False,,24576.0,False,False,True,torch.float32,,,False,10000.0,False,False
6
+ 20,https://huggingface.co/cybershiptrooper/InterpBench/tree/main/20,Detect spam messages based on appearance of spam keywords.,10,4,0.05,0.001,False,1.0,1.0,1.0,2000.0,gelu,0.1,,2.0,4.0,10.0,1.0,custom,4.0,16.0,gelu,14.0,1e-05,True,True,True,True,False,False,,False,,,,,,,gpt2,,cuda,1.0,causal,False,0.0,0.16,True,False,standard,False,2.0,False,,384.0,False,False,True,torch.float32,,,False,10000.0,False,False
7
+ 21,https://huggingface.co/cybershiptrooper/InterpBench/tree/main/21,Extract unique tokens from a string,10,4,0.05,0.01,False,1.0,1.0,0.4,500.0,gelu,1.0,,2.0,20.0,10.0,5.0,custom,4.0,80.0,gelu,5.0,1e-05,True,True,True,True,False,False,,False,,,,,,,gpt2,,cpu,1.0,causal,False,0.0,0.1885618083164127,True,False,standard,False,3.0,False,,9600.0,False,False,True,torch.float32,,,False,10000.0,False,False
8
+ 24,https://huggingface.co/cybershiptrooper/InterpBench/tree/main/24,Identifies the first occurrence of each token in a sequence.,10,4,0.05,0.01,False,1.0,1.0,0.4,500.0,gelu,1.0,,2.0,20.0,10.0,5.0,custom,4.0,80.0,gelu,5.0,1e-05,True,True,True,True,False,False,,False,,,,,,,gpt2,,cpu,1.0,causal,False,0.0,0.1885618083164127,True,False,standard,False,3.0,False,,9600.0,False,False,True,torch.float32,,,False,10000.0,False,False
9
+ 3,https://huggingface.co/cybershiptrooper/InterpBench/tree/main/3,Returns the fraction of 'x' in the input up to the i-th position for all i.,5,4,0.05,0.001,False,1.0,1.0,10.0,2000.0,gelu,0.1,,2.0,12.0,5.0,3.0,custom,4.0,48.0,gelu,6.0,1e-05,True,True,True,True,False,False,,False,,,,,,,gpt2,,cpu,1.0,causal,False,0.0,0.22188007849009167,True,False,standard,False,1.0,False,,3456.0,False,False,True,torch.float32,,,False,10000.0,False,False
10
+ 33,https://huggingface.co/cybershiptrooper/InterpBench/tree/main/33,Checks if each token's length is odd or even.,10,4,0.05,0.001,False,1.0,1.0,0.4,2000.0,gelu,0.1,,2.0,4.0,10.0,1.0,custom,4.0,16.0,gelu,10.0,1e-05,True,True,True,True,False,False,,False,,,,,,,gpt2,,cpu,1.0,causal,False,0.0,0.17457431218879393,True,False,standard,False,2.0,False,,384.0,False,False,True,torch.float32,,,False,10000.0,False,False
11
+ 34,https://huggingface.co/cybershiptrooper/InterpBench/tree/main/34,Calculate the ratio of vowels to consonants in each word.,10,4,0.05,0.001,False,1.0,1.0,0.4,2000.0,gelu,0.1,,2.0,4.0,10.0,1.0,custom,4.0,16.0,gelu,10.0,1e-05,True,True,True,True,False,False,,False,,,,,,,gpt2,,cpu,1.0,causal,False,0.0,0.16329931618554522,True,False,standard,False,5.0,False,,384.0,False,False,True,torch.float32,,,False,10000.0,False,False
12
+ 35,https://huggingface.co/cybershiptrooper/InterpBench/tree/main/35,Alternates capitalization of each character in words.,10,4,0.05,0.001,False,1.0,1.0,0.4,2000.0,gelu,0.1,,2.0,4.0,10.0,1.0,custom,4.0,16.0,gelu,10.0,1e-05,True,True,True,True,False,False,,False,,,,,,,gpt2,,cpu,1.0,causal,False,0.0,0.1539600717839002,True,False,standard,False,8.0,False,,384.0,False,False,True,torch.float32,,,False,10000.0,False,False
13
+ 36,https://huggingface.co/cybershiptrooper/InterpBench/tree/main/36,"Classifies each token as 'positive', 'negative', or 'neutral' based on emojis.",10,4,0.05,0.001,False,1.0,1.0,10.0,2000.0,gelu,0.1,,2.0,4.0,10.0,1.0,custom,4.0,16.0,gelu,5.0,1e-05,True,True,True,True,False,False,,False,,,,,,,gpt2,,cuda,1.0,causal,False,0.0,0.19402850002906638,True,False,standard,False,3.0,False,,384.0,False,False,True,torch.float32,,,False,10000.0,False,False
14
+ 37,https://huggingface.co/cybershiptrooper/InterpBench/tree/main/37,Reverses each word in the sequence except for specified exclusions.,10,4,0.05,0.001,False,1.0,1.0,0.4,2000.0,gelu,0.1,,2.0,4.0,10.0,1.0,custom,4.0,16.0,gelu,10.0,1e-05,True,True,True,True,False,False,,False,,,,,,,gpt2,,cpu,1.0,causal,False,0.0,0.1539600717839002,True,False,standard,False,8.0,False,,384.0,False,False,True,torch.float32,,,False,10000.0,False,False
15
+ 38,https://huggingface.co/cybershiptrooper/InterpBench/tree/main/38,Checks if tokens alternate between two types.,10,4,0.05,0.001,False,1.0,1.0,0.4,2000.0,gelu,0.1,,2.0,20.0,10.0,5.0,custom,4.0,80.0,gelu,5.0,1e-05,True,True,True,True,False,False,,False,,,,,,,gpt2,,cpu,1.0,causal,False,0.0,0.1539600717839002,True,False,standard,False,2.0,False,,9600.0,False,False,True,torch.float32,,,False,10000.0,False,False
16
+ 4,https://huggingface.co/cybershiptrooper/InterpBench/tree/main/4,Return fraction of previous open tokens minus the fraction of close tokens.,10,4,0.05,0.001,False,1.0,1.0,0.4,2000.0,gelu,0.1,,2.0,20.0,10.0,5.0,custom,4.0,80.0,gelu,7.0,1e-05,True,True,True,True,False,False,,False,,,,,,,gpt2,,cpu,1.0,causal,False,0.0,0.17056057308448835,True,False,standard,False,1.0,False,,9600.0,False,False,True,torch.float32,,,False,10000.0,False,False
17
+ 8,https://huggingface.co/cybershiptrooper/InterpBench/tree/main/8,Fills gaps between tokens with a specified filler.,10,4,0.05,0.01,False,1.0,1.0,0.4,500.0,gelu,1.0,,2.0,20.0,10.0,5.0,custom,4.0,80.0,gelu,10.0,1e-05,True,True,True,True,False,False,,False,,,,,,,gpt2,,cpu,1.0,causal,False,0.0,0.13333333333333333,True,False,standard,False,8.0,False,,9600.0,False,False,True,torch.float32,,,False,10000.0,False,False
18
+ ioi,https://huggingface.co/cybershiptrooper/InterpBench/tree/main/ioi,Indirect object identification,16,16,,,True,,,,,,,,,,,,,,,,,,True,True,True,True,True,True,,True,,,,,,,,,,,,True,,,True,True,,True,,True,,,True,True,True,,,,True,,True,True
19
+ ioi_next_token,https://huggingface.co/cybershiptrooper/InterpBench/tree/main/ioi_next_token,Indirect object identification,16,16,,,True,,,,,,,,,,,,,,,,,,True,True,True,True,True,True,,True,,,,,,,,,,,,True,,,True,True,,True,,True,,,True,True,True,,,,True,,True,True
benchmark_cases_metadata.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:01f473326ab84033784efd049808e3f0395bf577e23311a60f9b05eff32f5ca3
3
- size 49942
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:12652b82dbded2521f44e1219ade14c88e6bd787db5a2141803db257fb375e87
3
+ size 51034
benchmark_metadata.json CHANGED
@@ -2,28 +2,43 @@
2
  "name": "InterpBench",
3
  "version": "1.0.0",
4
  "description": "A benchmark of transformers with known circuits for evaluating mechanistic interpretability techniques.",
 
 
5
  "cases": [
6
  {
7
  "case_id": "11",
8
- "files": [
9
- "edges.pkl",
10
- "ll_model_510.pth",
11
- "ll_model_cfg_510.pkl",
12
- "meta_510.json"
13
- ],
14
  "task_description": "Counts the number of words in a sequence based on their length.",
15
  "vocab": [
16
  "J",
17
- "oCLrZaW",
18
- "no",
19
- "poiVg",
20
  "V",
21
  "b",
22
- "LB",
23
- "TPSI"
 
24
  ],
25
  "max_seq_len": 10,
26
  "min_seq_len": 4,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
27
  "transformer_cfg": {
28
  "n_layers": 2,
29
  "d_model": 12,
@@ -91,12 +106,7 @@
91
  },
92
  {
93
  "case_id": "13",
94
- "files": [
95
- "edges.pkl",
96
- "ll_model_510.pth",
97
- "ll_model_cfg_510.pkl",
98
- "meta_510.json"
99
- ],
100
  "task_description": "Analyzes the trend (increasing, decreasing, constant) of numeric tokens.",
101
  "vocab": [
102
  0,
@@ -105,6 +115,24 @@
105
  ],
106
  "max_seq_len": 10,
107
  "min_seq_len": 4,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
108
  "transformer_cfg": {
109
  "n_layers": 2,
110
  "d_model": 20,
@@ -172,22 +200,35 @@
172
  },
173
  {
174
  "case_id": "18",
175
- "files": [
176
- "edges.pkl",
177
- "ll_model_510.pth",
178
- "ll_model_cfg_510.pkl",
179
- "meta_510.json"
180
- ],
181
  "task_description": "Classify each token based on its frequency as 'rare', 'common', or 'frequent'.",
182
  "vocab": [
183
- "c",
184
- "e",
185
  "b",
 
186
  "d",
187
- "a"
188
  ],
189
  "max_seq_len": 10,
190
  "min_seq_len": 4,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
191
  "transformer_cfg": {
192
  "n_layers": 2,
193
  "d_model": 12,
@@ -255,20 +296,33 @@
255
  },
256
  {
257
  "case_id": "19",
258
- "files": [
259
- "edges.pkl",
260
- "ll_model_510.pth",
261
- "ll_model_cfg_510.pkl",
262
- "meta_510.json"
263
- ],
264
  "task_description": "Removes consecutive duplicate tokens from a sequence.",
265
  "vocab": [
266
- "b",
267
  "a",
 
268
  "c"
269
  ],
270
  "max_seq_len": 15,
271
  "min_seq_len": 4,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
272
  "transformer_cfg": {
273
  "n_layers": 2,
274
  "d_model": 32,
@@ -336,29 +390,42 @@
336
  },
337
  {
338
  "case_id": "20",
339
- "files": [
340
- "edges.pkl",
341
- "ll_model_1110.pth",
342
- "ll_model_cfg_1110.pkl",
343
- "meta_1110.json"
344
- ],
345
  "task_description": "Detect spam messages based on appearance of spam keywords.",
346
  "vocab": [
347
  "J",
348
- "spam",
349
- "offer",
350
- "click",
351
- "oCLrZaW",
352
- "no",
353
- "poiVg",
354
  "V",
355
  "b",
356
- "LB",
 
357
  "now",
358
- "TPSI"
 
 
 
359
  ],
360
  "max_seq_len": 10,
361
  "min_seq_len": 4,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
362
  "transformer_cfg": {
363
  "n_layers": 2,
364
  "d_model": 4,
@@ -426,20 +493,33 @@
426
  },
427
  {
428
  "case_id": "21",
429
- "files": [
430
- "edges.pkl",
431
- "ll_model_510.pth",
432
- "ll_model_cfg_510.pkl",
433
- "meta_510.json"
434
- ],
435
  "task_description": "Extract unique tokens from a string",
436
  "vocab": [
437
- "b",
438
  "a",
 
439
  "c"
440
  ],
441
  "max_seq_len": 10,
442
  "min_seq_len": 4,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
443
  "transformer_cfg": {
444
  "n_layers": 2,
445
  "d_model": 20,
@@ -507,20 +587,33 @@
507
  },
508
  {
509
  "case_id": "24",
510
- "files": [
511
- "edges.pkl",
512
- "ll_model_510.pth",
513
- "ll_model_cfg_510.pkl",
514
- "meta_510.json"
515
- ],
516
  "task_description": "Identifies the first occurrence of each token in a sequence.",
517
  "vocab": [
518
- "b",
519
  "a",
 
520
  "c"
521
  ],
522
  "max_seq_len": 10,
523
  "min_seq_len": 4,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
524
  "transformer_cfg": {
525
  "n_layers": 2,
526
  "d_model": 20,
@@ -588,21 +681,34 @@
588
  },
589
  {
590
  "case_id": "3",
591
- "files": [
592
- "edges.pkl",
593
- "ll_model_10110.pth",
594
- "ll_model_cfg_10110.pkl",
595
- "meta_10110.json"
596
- ],
597
  "task_description": "Returns the fraction of 'x' in the input up to the i-th position for all i.",
598
  "vocab": [
599
- "x",
600
- "b",
601
  "a",
602
- "c"
 
 
603
  ],
604
  "max_seq_len": 5,
605
  "min_seq_len": 4,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
606
  "transformer_cfg": {
607
  "n_layers": 2,
608
  "d_model": 12,
@@ -670,25 +776,38 @@
670
  },
671
  {
672
  "case_id": "33",
673
- "files": [
674
- "edges.pkl",
675
- "ll_model_510.pth",
676
- "ll_model_cfg_510.pkl",
677
- "meta_510.json"
678
- ],
679
  "task_description": "Checks if each token's length is odd or even.",
680
  "vocab": [
681
  "J",
682
- "oCLrZaW",
683
- "no",
684
- "poiVg",
685
  "V",
686
  "b",
687
- "LB",
688
- "TPSI"
 
689
  ],
690
  "max_seq_len": 10,
691
  "min_seq_len": 4,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
692
  "transformer_cfg": {
693
  "n_layers": 2,
694
  "d_model": 4,
@@ -756,25 +875,38 @@
756
  },
757
  {
758
  "case_id": "34",
759
- "files": [
760
- "edges.pkl",
761
- "ll_model_510.pth",
762
- "ll_model_cfg_510.pkl",
763
- "meta_510.json"
764
- ],
765
  "task_description": "Calculate the ratio of vowels to consonants in each word.",
766
  "vocab": [
767
  "J",
768
- "oCLrZaW",
769
- "no",
770
- "poiVg",
771
  "V",
772
  "b",
773
- "LB",
774
- "TPSI"
 
775
  ],
776
  "max_seq_len": 10,
777
  "min_seq_len": 4,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
778
  "transformer_cfg": {
779
  "n_layers": 2,
780
  "d_model": 4,
@@ -842,25 +974,38 @@
842
  },
843
  {
844
  "case_id": "35",
845
- "files": [
846
- "edges.pkl",
847
- "ll_model_510.pth",
848
- "ll_model_cfg_510.pkl",
849
- "meta_510.json"
850
- ],
851
  "task_description": "Alternates capitalization of each character in words.",
852
  "vocab": [
853
  "J",
854
- "oCLrZaW",
855
- "no",
856
- "poiVg",
857
  "V",
858
  "b",
859
- "LB",
860
- "TPSI"
 
861
  ],
862
  "max_seq_len": 10,
863
  "min_seq_len": 4,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
864
  "transformer_cfg": {
865
  "n_layers": 2,
866
  "d_model": 4,
@@ -928,20 +1073,33 @@
928
  },
929
  {
930
  "case_id": "36",
931
- "files": [
932
- "edges.pkl",
933
- "ll_model_10110.pth",
934
- "ll_model_cfg_10110.pkl",
935
- "meta_10110.json"
936
- ],
937
  "task_description": "Classifies each token as 'positive', 'negative', or 'neutral' based on emojis.",
938
  "vocab": [
939
- "\ud83d\ude22",
940
  "\ud83d\udcd8",
941
- "\ud83d\ude0a"
 
942
  ],
943
  "max_seq_len": 10,
944
  "min_seq_len": 4,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
945
  "transformer_cfg": {
946
  "n_layers": 2,
947
  "d_model": 4,
@@ -1009,25 +1167,38 @@
1009
  },
1010
  {
1011
  "case_id": "37",
1012
- "files": [
1013
- "edges.pkl",
1014
- "ll_model_510.pth",
1015
- "ll_model_cfg_510.pkl",
1016
- "meta_510.json"
1017
- ],
1018
  "task_description": "Reverses each word in the sequence except for specified exclusions.",
1019
  "vocab": [
1020
  "J",
1021
- "oCLrZaW",
1022
- "no",
1023
- "poiVg",
1024
  "V",
1025
  "b",
1026
- "LB",
1027
- "TPSI"
 
1028
  ],
1029
  "max_seq_len": 10,
1030
  "min_seq_len": 4,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1031
  "transformer_cfg": {
1032
  "n_layers": 2,
1033
  "d_model": 4,
@@ -1095,20 +1266,33 @@
1095
  },
1096
  {
1097
  "case_id": "38",
1098
- "files": [
1099
- "edges.pkl",
1100
- "ll_model_510.pth",
1101
- "ll_model_cfg_510.pkl",
1102
- "meta_510.json"
1103
- ],
1104
  "task_description": "Checks if tokens alternate between two types.",
1105
  "vocab": [
1106
- "b",
1107
  "a",
 
1108
  "c"
1109
  ],
1110
  "max_seq_len": 10,
1111
  "min_seq_len": 4,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1112
  "transformer_cfg": {
1113
  "n_layers": 2,
1114
  "d_model": 20,
@@ -1176,22 +1360,35 @@
1176
  },
1177
  {
1178
  "case_id": "4",
1179
- "files": [
1180
- "edges.pkl",
1181
- "ll_model_510.pth",
1182
- "ll_model_cfg_510.pkl",
1183
- "meta_510.json"
1184
- ],
1185
  "task_description": "Return fraction of previous open tokens minus the fraction of close tokens.",
1186
  "vocab": [
1187
- "b",
1188
  "(",
1189
- "c",
1190
  ")",
1191
- "a"
 
 
1192
  ],
1193
  "max_seq_len": 10,
1194
  "min_seq_len": 4,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1195
  "transformer_cfg": {
1196
  "n_layers": 2,
1197
  "d_model": 20,
@@ -1259,25 +1456,38 @@
1259
  },
1260
  {
1261
  "case_id": "8",
1262
- "files": [
1263
- "edges.pkl",
1264
- "ll_model_510.pth",
1265
- "ll_model_cfg_510.pkl",
1266
- "meta_510.json"
1267
- ],
1268
  "task_description": "Fills gaps between tokens with a specified filler.",
1269
  "vocab": [
1270
  "J",
1271
- "oCLrZaW",
1272
- "no",
1273
- "poiVg",
1274
  "V",
1275
  "b",
1276
- "LB",
1277
- "TPSI"
 
1278
  ],
1279
  "max_seq_len": 10,
1280
  "min_seq_len": 4,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1281
  "transformer_cfg": {
1282
  "n_layers": 2,
1283
  "d_model": 20,
@@ -1345,17 +1555,40 @@
1345
  },
1346
  {
1347
  "case_id": "ioi",
 
 
 
 
1348
  "files": [
1349
- "corr_100_100_40.json",
1350
- "ll_model_100_100_40.pth"
 
 
 
 
 
 
1351
  ]
1352
  },
1353
  {
1354
  "case_id": "ioi_next_token",
 
 
 
 
1355
  "files": [
1356
- "corr_100_100_40.json",
1357
- "ll_model_100_100_40.pth",
1358
- "training_args.json"
 
 
 
 
 
 
 
 
 
1359
  ]
1360
  }
1361
  ]
 
2
  "name": "InterpBench",
3
  "version": "1.0.0",
4
  "description": "A benchmark of transformers with known circuits for evaluating mechanistic interpretability techniques.",
5
+ "license": "https://creativecommons.org/licenses/by/4.0/",
6
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench",
7
  "cases": [
8
  {
9
  "case_id": "11",
10
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/tree/main/11",
 
 
 
 
 
11
  "task_description": "Counts the number of words in a sequence based on their length.",
12
  "vocab": [
13
  "J",
14
+ "LB",
15
+ "TPSI",
 
16
  "V",
17
  "b",
18
+ "no",
19
+ "oCLrZaW",
20
+ "poiVg"
21
  ],
22
  "max_seq_len": 10,
23
  "min_seq_len": 4,
24
+ "files": [
25
+ {
26
+ "file_name": "edges.pkl",
27
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/11/edges.pkl"
28
+ },
29
+ {
30
+ "file_name": "ll_model_510.pth",
31
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/11/ll_model_510.pth"
32
+ },
33
+ {
34
+ "file_name": "ll_model_cfg_510.pkl",
35
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/11/ll_model_cfg_510.pkl"
36
+ },
37
+ {
38
+ "file_name": "meta_510.json",
39
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/11/meta_510.json"
40
+ }
41
+ ],
42
  "transformer_cfg": {
43
  "n_layers": 2,
44
  "d_model": 12,
 
106
  },
107
  {
108
  "case_id": "13",
109
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/tree/main/13",
 
 
 
 
 
110
  "task_description": "Analyzes the trend (increasing, decreasing, constant) of numeric tokens.",
111
  "vocab": [
112
  0,
 
115
  ],
116
  "max_seq_len": 10,
117
  "min_seq_len": 4,
118
+ "files": [
119
+ {
120
+ "file_name": "edges.pkl",
121
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/13/edges.pkl"
122
+ },
123
+ {
124
+ "file_name": "ll_model_510.pth",
125
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/13/ll_model_510.pth"
126
+ },
127
+ {
128
+ "file_name": "ll_model_cfg_510.pkl",
129
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/13/ll_model_cfg_510.pkl"
130
+ },
131
+ {
132
+ "file_name": "meta_510.json",
133
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/13/meta_510.json"
134
+ }
135
+ ],
136
  "transformer_cfg": {
137
  "n_layers": 2,
138
  "d_model": 20,
 
200
  },
201
  {
202
  "case_id": "18",
203
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/tree/main/18",
 
 
 
 
 
204
  "task_description": "Classify each token based on its frequency as 'rare', 'common', or 'frequent'.",
205
  "vocab": [
206
+ "a",
 
207
  "b",
208
+ "c",
209
  "d",
210
+ "e"
211
  ],
212
  "max_seq_len": 10,
213
  "min_seq_len": 4,
214
+ "files": [
215
+ {
216
+ "file_name": "edges.pkl",
217
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/18/edges.pkl"
218
+ },
219
+ {
220
+ "file_name": "ll_model_510.pth",
221
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/18/ll_model_510.pth"
222
+ },
223
+ {
224
+ "file_name": "ll_model_cfg_510.pkl",
225
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/18/ll_model_cfg_510.pkl"
226
+ },
227
+ {
228
+ "file_name": "meta_510.json",
229
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/18/meta_510.json"
230
+ }
231
+ ],
232
  "transformer_cfg": {
233
  "n_layers": 2,
234
  "d_model": 12,
 
296
  },
297
  {
298
  "case_id": "19",
299
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/tree/main/19",
 
 
 
 
 
300
  "task_description": "Removes consecutive duplicate tokens from a sequence.",
301
  "vocab": [
 
302
  "a",
303
+ "b",
304
  "c"
305
  ],
306
  "max_seq_len": 15,
307
  "min_seq_len": 4,
308
+ "files": [
309
+ {
310
+ "file_name": "edges.pkl",
311
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/19/edges.pkl"
312
+ },
313
+ {
314
+ "file_name": "ll_model_510.pth",
315
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/19/ll_model_510.pth"
316
+ },
317
+ {
318
+ "file_name": "ll_model_cfg_510.pkl",
319
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/19/ll_model_cfg_510.pkl"
320
+ },
321
+ {
322
+ "file_name": "meta_510.json",
323
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/19/meta_510.json"
324
+ }
325
+ ],
326
  "transformer_cfg": {
327
  "n_layers": 2,
328
  "d_model": 32,
 
390
  },
391
  {
392
  "case_id": "20",
393
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/tree/main/20",
 
 
 
 
 
394
  "task_description": "Detect spam messages based on appearance of spam keywords.",
395
  "vocab": [
396
  "J",
397
+ "LB",
398
+ "TPSI",
 
 
 
 
399
  "V",
400
  "b",
401
+ "click",
402
+ "no",
403
  "now",
404
+ "oCLrZaW",
405
+ "offer",
406
+ "poiVg",
407
+ "spam"
408
  ],
409
  "max_seq_len": 10,
410
  "min_seq_len": 4,
411
+ "files": [
412
+ {
413
+ "file_name": "edges.pkl",
414
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/20/edges.pkl"
415
+ },
416
+ {
417
+ "file_name": "ll_model_1110.pth",
418
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/20/ll_model_1110.pth"
419
+ },
420
+ {
421
+ "file_name": "ll_model_cfg_1110.pkl",
422
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/20/ll_model_cfg_1110.pkl"
423
+ },
424
+ {
425
+ "file_name": "meta_1110.json",
426
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/20/meta_1110.json"
427
+ }
428
+ ],
429
  "transformer_cfg": {
430
  "n_layers": 2,
431
  "d_model": 4,
 
493
  },
494
  {
495
  "case_id": "21",
496
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/tree/main/21",
 
 
 
 
 
497
  "task_description": "Extract unique tokens from a string",
498
  "vocab": [
 
499
  "a",
500
+ "b",
501
  "c"
502
  ],
503
  "max_seq_len": 10,
504
  "min_seq_len": 4,
505
+ "files": [
506
+ {
507
+ "file_name": "edges.pkl",
508
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/21/edges.pkl"
509
+ },
510
+ {
511
+ "file_name": "ll_model_510.pth",
512
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/21/ll_model_510.pth"
513
+ },
514
+ {
515
+ "file_name": "ll_model_cfg_510.pkl",
516
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/21/ll_model_cfg_510.pkl"
517
+ },
518
+ {
519
+ "file_name": "meta_510.json",
520
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/21/meta_510.json"
521
+ }
522
+ ],
523
  "transformer_cfg": {
524
  "n_layers": 2,
525
  "d_model": 20,
 
587
  },
588
  {
589
  "case_id": "24",
590
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/tree/main/24",
 
 
 
 
 
591
  "task_description": "Identifies the first occurrence of each token in a sequence.",
592
  "vocab": [
 
593
  "a",
594
+ "b",
595
  "c"
596
  ],
597
  "max_seq_len": 10,
598
  "min_seq_len": 4,
599
+ "files": [
600
+ {
601
+ "file_name": "edges.pkl",
602
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/24/edges.pkl"
603
+ },
604
+ {
605
+ "file_name": "ll_model_510.pth",
606
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/24/ll_model_510.pth"
607
+ },
608
+ {
609
+ "file_name": "ll_model_cfg_510.pkl",
610
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/24/ll_model_cfg_510.pkl"
611
+ },
612
+ {
613
+ "file_name": "meta_510.json",
614
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/24/meta_510.json"
615
+ }
616
+ ],
617
  "transformer_cfg": {
618
  "n_layers": 2,
619
  "d_model": 20,
 
681
  },
682
  {
683
  "case_id": "3",
684
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/tree/main/3",
 
 
 
 
 
685
  "task_description": "Returns the fraction of 'x' in the input up to the i-th position for all i.",
686
  "vocab": [
 
 
687
  "a",
688
+ "b",
689
+ "c",
690
+ "x"
691
  ],
692
  "max_seq_len": 5,
693
  "min_seq_len": 4,
694
+ "files": [
695
+ {
696
+ "file_name": "edges.pkl",
697
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/3/edges.pkl"
698
+ },
699
+ {
700
+ "file_name": "ll_model_10110.pth",
701
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/3/ll_model_10110.pth"
702
+ },
703
+ {
704
+ "file_name": "ll_model_cfg_10110.pkl",
705
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/3/ll_model_cfg_10110.pkl"
706
+ },
707
+ {
708
+ "file_name": "meta_10110.json",
709
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/3/meta_10110.json"
710
+ }
711
+ ],
712
  "transformer_cfg": {
713
  "n_layers": 2,
714
  "d_model": 12,
 
776
  },
777
  {
778
  "case_id": "33",
779
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/tree/main/33",
 
 
 
 
 
780
  "task_description": "Checks if each token's length is odd or even.",
781
  "vocab": [
782
  "J",
783
+ "LB",
784
+ "TPSI",
 
785
  "V",
786
  "b",
787
+ "no",
788
+ "oCLrZaW",
789
+ "poiVg"
790
  ],
791
  "max_seq_len": 10,
792
  "min_seq_len": 4,
793
+ "files": [
794
+ {
795
+ "file_name": "edges.pkl",
796
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/33/edges.pkl"
797
+ },
798
+ {
799
+ "file_name": "ll_model_510.pth",
800
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/33/ll_model_510.pth"
801
+ },
802
+ {
803
+ "file_name": "ll_model_cfg_510.pkl",
804
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/33/ll_model_cfg_510.pkl"
805
+ },
806
+ {
807
+ "file_name": "meta_510.json",
808
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/33/meta_510.json"
809
+ }
810
+ ],
811
  "transformer_cfg": {
812
  "n_layers": 2,
813
  "d_model": 4,
 
875
  },
876
  {
877
  "case_id": "34",
878
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/tree/main/34",
 
 
 
 
 
879
  "task_description": "Calculate the ratio of vowels to consonants in each word.",
880
  "vocab": [
881
  "J",
882
+ "LB",
883
+ "TPSI",
 
884
  "V",
885
  "b",
886
+ "no",
887
+ "oCLrZaW",
888
+ "poiVg"
889
  ],
890
  "max_seq_len": 10,
891
  "min_seq_len": 4,
892
+ "files": [
893
+ {
894
+ "file_name": "edges.pkl",
895
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/34/edges.pkl"
896
+ },
897
+ {
898
+ "file_name": "ll_model_510.pth",
899
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/34/ll_model_510.pth"
900
+ },
901
+ {
902
+ "file_name": "ll_model_cfg_510.pkl",
903
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/34/ll_model_cfg_510.pkl"
904
+ },
905
+ {
906
+ "file_name": "meta_510.json",
907
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/34/meta_510.json"
908
+ }
909
+ ],
910
  "transformer_cfg": {
911
  "n_layers": 2,
912
  "d_model": 4,
 
974
  },
975
  {
976
  "case_id": "35",
977
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/tree/main/35",
 
 
 
 
 
978
  "task_description": "Alternates capitalization of each character in words.",
979
  "vocab": [
980
  "J",
981
+ "LB",
982
+ "TPSI",
 
983
  "V",
984
  "b",
985
+ "no",
986
+ "oCLrZaW",
987
+ "poiVg"
988
  ],
989
  "max_seq_len": 10,
990
  "min_seq_len": 4,
991
+ "files": [
992
+ {
993
+ "file_name": "edges.pkl",
994
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/35/edges.pkl"
995
+ },
996
+ {
997
+ "file_name": "ll_model_510.pth",
998
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/35/ll_model_510.pth"
999
+ },
1000
+ {
1001
+ "file_name": "ll_model_cfg_510.pkl",
1002
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/35/ll_model_cfg_510.pkl"
1003
+ },
1004
+ {
1005
+ "file_name": "meta_510.json",
1006
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/35/meta_510.json"
1007
+ }
1008
+ ],
1009
  "transformer_cfg": {
1010
  "n_layers": 2,
1011
  "d_model": 4,
 
1073
  },
1074
  {
1075
  "case_id": "36",
1076
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/tree/main/36",
 
 
 
 
 
1077
  "task_description": "Classifies each token as 'positive', 'negative', or 'neutral' based on emojis.",
1078
  "vocab": [
 
1079
  "\ud83d\udcd8",
1080
+ "\ud83d\ude0a",
1081
+ "\ud83d\ude22"
1082
  ],
1083
  "max_seq_len": 10,
1084
  "min_seq_len": 4,
1085
+ "files": [
1086
+ {
1087
+ "file_name": "edges.pkl",
1088
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/36/edges.pkl"
1089
+ },
1090
+ {
1091
+ "file_name": "ll_model_10110.pth",
1092
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/36/ll_model_10110.pth"
1093
+ },
1094
+ {
1095
+ "file_name": "ll_model_cfg_10110.pkl",
1096
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/36/ll_model_cfg_10110.pkl"
1097
+ },
1098
+ {
1099
+ "file_name": "meta_10110.json",
1100
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/36/meta_10110.json"
1101
+ }
1102
+ ],
1103
  "transformer_cfg": {
1104
  "n_layers": 2,
1105
  "d_model": 4,
 
1167
  },
1168
  {
1169
  "case_id": "37",
1170
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/tree/main/37",
 
 
 
 
 
1171
  "task_description": "Reverses each word in the sequence except for specified exclusions.",
1172
  "vocab": [
1173
  "J",
1174
+ "LB",
1175
+ "TPSI",
 
1176
  "V",
1177
  "b",
1178
+ "no",
1179
+ "oCLrZaW",
1180
+ "poiVg"
1181
  ],
1182
  "max_seq_len": 10,
1183
  "min_seq_len": 4,
1184
+ "files": [
1185
+ {
1186
+ "file_name": "edges.pkl",
1187
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/37/edges.pkl"
1188
+ },
1189
+ {
1190
+ "file_name": "ll_model_510.pth",
1191
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/37/ll_model_510.pth"
1192
+ },
1193
+ {
1194
+ "file_name": "ll_model_cfg_510.pkl",
1195
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/37/ll_model_cfg_510.pkl"
1196
+ },
1197
+ {
1198
+ "file_name": "meta_510.json",
1199
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/37/meta_510.json"
1200
+ }
1201
+ ],
1202
  "transformer_cfg": {
1203
  "n_layers": 2,
1204
  "d_model": 4,
 
1266
  },
1267
  {
1268
  "case_id": "38",
1269
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/tree/main/38",
 
 
 
 
 
1270
  "task_description": "Checks if tokens alternate between two types.",
1271
  "vocab": [
 
1272
  "a",
1273
+ "b",
1274
  "c"
1275
  ],
1276
  "max_seq_len": 10,
1277
  "min_seq_len": 4,
1278
+ "files": [
1279
+ {
1280
+ "file_name": "edges.pkl",
1281
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/38/edges.pkl"
1282
+ },
1283
+ {
1284
+ "file_name": "ll_model_510.pth",
1285
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/38/ll_model_510.pth"
1286
+ },
1287
+ {
1288
+ "file_name": "ll_model_cfg_510.pkl",
1289
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/38/ll_model_cfg_510.pkl"
1290
+ },
1291
+ {
1292
+ "file_name": "meta_510.json",
1293
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/38/meta_510.json"
1294
+ }
1295
+ ],
1296
  "transformer_cfg": {
1297
  "n_layers": 2,
1298
  "d_model": 20,
 
1360
  },
1361
  {
1362
  "case_id": "4",
1363
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/tree/main/4",
 
 
 
 
 
1364
  "task_description": "Return fraction of previous open tokens minus the fraction of close tokens.",
1365
  "vocab": [
 
1366
  "(",
 
1367
  ")",
1368
+ "a",
1369
+ "b",
1370
+ "c"
1371
  ],
1372
  "max_seq_len": 10,
1373
  "min_seq_len": 4,
1374
+ "files": [
1375
+ {
1376
+ "file_name": "edges.pkl",
1377
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/4/edges.pkl"
1378
+ },
1379
+ {
1380
+ "file_name": "ll_model_510.pth",
1381
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/4/ll_model_510.pth"
1382
+ },
1383
+ {
1384
+ "file_name": "ll_model_cfg_510.pkl",
1385
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/4/ll_model_cfg_510.pkl"
1386
+ },
1387
+ {
1388
+ "file_name": "meta_510.json",
1389
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/4/meta_510.json"
1390
+ }
1391
+ ],
1392
  "transformer_cfg": {
1393
  "n_layers": 2,
1394
  "d_model": 20,
 
1456
  },
1457
  {
1458
  "case_id": "8",
1459
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/tree/main/8",
 
 
 
 
 
1460
  "task_description": "Fills gaps between tokens with a specified filler.",
1461
  "vocab": [
1462
  "J",
1463
+ "LB",
1464
+ "TPSI",
 
1465
  "V",
1466
  "b",
1467
+ "no",
1468
+ "oCLrZaW",
1469
+ "poiVg"
1470
  ],
1471
  "max_seq_len": 10,
1472
  "min_seq_len": 4,
1473
+ "files": [
1474
+ {
1475
+ "file_name": "edges.pkl",
1476
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/8/edges.pkl"
1477
+ },
1478
+ {
1479
+ "file_name": "ll_model_510.pth",
1480
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/8/ll_model_510.pth"
1481
+ },
1482
+ {
1483
+ "file_name": "ll_model_cfg_510.pkl",
1484
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/8/ll_model_cfg_510.pkl"
1485
+ },
1486
+ {
1487
+ "file_name": "meta_510.json",
1488
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/8/meta_510.json"
1489
+ }
1490
+ ],
1491
  "transformer_cfg": {
1492
  "n_layers": 2,
1493
  "d_model": 20,
 
1555
  },
1556
  {
1557
  "case_id": "ioi",
1558
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/tree/main/ioi",
1559
+ "task_description": "Indirect object identification",
1560
+ "max_seq_len": 16,
1561
+ "min_seq_len": 16,
1562
  "files": [
1563
+ {
1564
+ "file_name": "corr_100_100_40.json",
1565
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/ioi/corr_100_100_40.json"
1566
+ },
1567
+ {
1568
+ "file_name": "ll_model_100_100_40.pth",
1569
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/ioi/ll_model_100_100_40.pth"
1570
+ }
1571
  ]
1572
  },
1573
  {
1574
  "case_id": "ioi_next_token",
1575
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/tree/main/ioi_next_token",
1576
+ "task_description": "Indirect object identification",
1577
+ "max_seq_len": 16,
1578
+ "min_seq_len": 16,
1579
  "files": [
1580
+ {
1581
+ "file_name": "corr_100_100_40.json",
1582
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/ioi_next_token/corr_100_100_40.json"
1583
+ },
1584
+ {
1585
+ "file_name": "ll_model_100_100_40.pth",
1586
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/ioi_next_token/ll_model_100_100_40.pth"
1587
+ },
1588
+ {
1589
+ "file_name": "training_args.json",
1590
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench/blob/main/ioi_next_token/training_args.json"
1591
+ }
1592
  ]
1593
  }
1594
  ]
benchmark_metadata_croissant.json ADDED
@@ -0,0 +1,1103 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "@context": {
3
+ "@language": "en",
4
+ "@vocab": "https://schema.org/",
5
+ "citeAs": "cr:citeAs",
6
+ "column": "cr:column",
7
+ "conformsTo": "dct:conformsTo",
8
+ "cr": "http://mlcommons.org/croissant/",
9
+ "rai": "http://mlcommons.org/croissant/RAI/",
10
+ "data": {
11
+ "@id": "cr:data",
12
+ "@type": "@json"
13
+ },
14
+ "dataType": {
15
+ "@id": "cr:dataType",
16
+ "@type": "@vocab"
17
+ },
18
+ "dct": "http://purl.org/dc/terms/",
19
+ "examples": {
20
+ "@id": "cr:examples",
21
+ "@type": "@json"
22
+ },
23
+ "extract": "cr:extract",
24
+ "field": "cr:field",
25
+ "fileProperty": "cr:fileProperty",
26
+ "fileObject": "cr:fileObject",
27
+ "fileSet": "cr:fileSet",
28
+ "format": "cr:format",
29
+ "includes": "cr:includes",
30
+ "isLiveDataset": "cr:isLiveDataset",
31
+ "jsonPath": "cr:jsonPath",
32
+ "key": "cr:key",
33
+ "md5": "cr:md5",
34
+ "parentField": "cr:parentField",
35
+ "path": "cr:path",
36
+ "recordSet": "cr:recordSet",
37
+ "references": "cr:references",
38
+ "regex": "cr:regex",
39
+ "repeated": "cr:repeated",
40
+ "replace": "cr:replace",
41
+ "sc": "https://schema.org/",
42
+ "separator": "cr:separator",
43
+ "source": "cr:source",
44
+ "subField": "cr:subField",
45
+ "transform": "cr:transform"
46
+ },
47
+ "@type": "sc:Dataset",
48
+ "name": "InterpBench",
49
+ "description": "A benchmark of transformers with known circuits for evaluating mechanistic interpretability techniques.",
50
+ "conformsTo": "http://mlcommons.org/croissant/1.0",
51
+ "license": "https://creativecommons.org/licenses/by/4.0/",
52
+ "url": "https://huggingface.co/cybershiptrooper/InterpBench",
53
+ "version": "1.0.0",
54
+ "distribution": [
55
+ {
56
+ "@type": "cr:FileObject",
57
+ "@id": "hf-repository",
58
+ "name": "hf-repository",
59
+ "description": "The Hugging Face git repository.",
60
+ "contentUrl": "https://huggingface.co/cybershiptrooper/InterpBench",
61
+ "encodingFormat": "git+https",
62
+ "sha256": "main"
63
+ },
64
+ {
65
+ "@type": "cr:FileObject",
66
+ "@id": "benchmark-cases-parquet",
67
+ "name": "benchmark-cases-parquet",
68
+ "description": "Parquet file describing all the cases in the benchmark.",
69
+ "containedIn": {
70
+ "@id": "hf-repository"
71
+ },
72
+ "encodingFormat": "application/x-parquet"
73
+ },
74
+ {
75
+ "@type": "cr:FileSet",
76
+ "@id": "training-args",
77
+ "name": "training-args",
78
+ "description": "Training arguments.",
79
+ "containedIn": {
80
+ "@id": "hf-repository"
81
+ },
82
+ "encodingFormat": "application/json",
83
+ "includes": "*/meta_[0-9]*.json"
84
+ },
85
+ {
86
+ "@type": "cr:FileSet",
87
+ "@id": "circuits",
88
+ "name": "circuits",
89
+ "description": "Ground truth circuits (Pickle).",
90
+ "containedIn": {
91
+ "@id": "hf-repository"
92
+ },
93
+ "encodingFormat": "application/octet-stream",
94
+ "includes": "*/edges.pkl"
95
+ },
96
+ {
97
+ "@type": "cr:FileSet",
98
+ "@id": "weights",
99
+ "name": "weights",
100
+ "description": "Serialized PyTorch state dictionaries (Pickle).",
101
+ "containedIn": {
102
+ "@id": "hf-repository"
103
+ },
104
+ "encodingFormat": "application/octet-stream",
105
+ "includes": "*/ll_model_[0-9]*.pkl"
106
+ },
107
+ {
108
+ "@type": "cr:FileSet",
109
+ "@id": "cfgs",
110
+ "name": "cfgs",
111
+ "description": "Architecture configs (Pickle).",
112
+ "containedIn": {
113
+ "@id": "hf-repository"
114
+ },
115
+ "encodingFormat": "application/octet-stream",
116
+ "includes": "*/ll_model_cfg_[0-9]*.pkl"
117
+ }
118
+ ],
119
+ "recordSet": [
120
+ {
121
+ "@type": "cr:RecordSet",
122
+ "@id": "cases",
123
+ "name": "cases",
124
+ "field": [
125
+ {
126
+ "@type": "cr:Field",
127
+ "@id": "case_id",
128
+ "name": "case_id",
129
+ "description": "Column 'case_id' from the parquet file describing all the cases in the benchmark.",
130
+ "dataType": "sc:Text",
131
+ "source": {
132
+ "fileSet": {
133
+ "@id": "benchmark-cases-parquet"
134
+ },
135
+ "extract": {
136
+ "column": "case_id"
137
+ }
138
+ }
139
+ },
140
+ {
141
+ "@type": "cr:Field",
142
+ "@id": "url",
143
+ "name": "url",
144
+ "description": "Column 'url' from the parquet file describing all the cases in the benchmark.",
145
+ "dataType": "sc:Text",
146
+ "source": {
147
+ "fileSet": {
148
+ "@id": "benchmark-cases-parquet"
149
+ },
150
+ "extract": {
151
+ "column": "url"
152
+ }
153
+ }
154
+ },
155
+ {
156
+ "@type": "cr:Field",
157
+ "@id": "task_description",
158
+ "name": "task_description",
159
+ "description": "Column 'task_description' from the parquet file describing all the cases in the benchmark.",
160
+ "dataType": "sc:Text",
161
+ "source": {
162
+ "fileSet": {
163
+ "@id": "benchmark-cases-parquet"
164
+ },
165
+ "extract": {
166
+ "column": "task_description"
167
+ }
168
+ }
169
+ },
170
+ {
171
+ "@type": "cr:Field",
172
+ "@id": "max_seq_len",
173
+ "name": "max_seq_len",
174
+ "description": "Column 'max_seq_len' from the parquet file describing all the cases in the benchmark.",
175
+ "dataType": "sc:Integer",
176
+ "source": {
177
+ "fileSet": {
178
+ "@id": "benchmark-cases-parquet"
179
+ },
180
+ "extract": {
181
+ "column": "max_seq_len"
182
+ }
183
+ }
184
+ },
185
+ {
186
+ "@type": "cr:Field",
187
+ "@id": "min_seq_len",
188
+ "name": "min_seq_len",
189
+ "description": "Column 'min_seq_len' from the parquet file describing all the cases in the benchmark.",
190
+ "dataType": "sc:Integer",
191
+ "source": {
192
+ "fileSet": {
193
+ "@id": "benchmark-cases-parquet"
194
+ },
195
+ "extract": {
196
+ "column": "min_seq_len"
197
+ }
198
+ }
199
+ },
200
+ {
201
+ "@type": "cr:Field",
202
+ "@id": "training_args.atol",
203
+ "name": "training_args.atol",
204
+ "description": "Column 'training_args.atol' from the parquet file describing all the cases in the benchmark.",
205
+ "dataType": "sc:Float",
206
+ "source": {
207
+ "fileSet": {
208
+ "@id": "benchmark-cases-parquet"
209
+ },
210
+ "extract": {
211
+ "column": "training_args.atol"
212
+ }
213
+ }
214
+ },
215
+ {
216
+ "@type": "cr:Field",
217
+ "@id": "training_args.lr",
218
+ "name": "training_args.lr",
219
+ "description": "Column 'training_args.lr' from the parquet file describing all the cases in the benchmark.",
220
+ "dataType": "sc:Float",
221
+ "source": {
222
+ "fileSet": {
223
+ "@id": "benchmark-cases-parquet"
224
+ },
225
+ "extract": {
226
+ "column": "training_args.lr"
227
+ }
228
+ }
229
+ },
230
+ {
231
+ "@type": "cr:Field",
232
+ "@id": "training_args.use_single_loss",
233
+ "name": "training_args.use_single_loss",
234
+ "description": "Column 'training_args.use_single_loss' from the parquet file describing all the cases in the benchmark.",
235
+ "dataType": "sc:Boolean",
236
+ "source": {
237
+ "fileSet": {
238
+ "@id": "benchmark-cases-parquet"
239
+ },
240
+ "extract": {
241
+ "column": "training_args.use_single_loss"
242
+ }
243
+ }
244
+ },
245
+ {
246
+ "@type": "cr:Field",
247
+ "@id": "training_args.iit_weight",
248
+ "name": "training_args.iit_weight",
249
+ "description": "Column 'training_args.iit_weight' from the parquet file describing all the cases in the benchmark.",
250
+ "dataType": "sc:Float",
251
+ "source": {
252
+ "fileSet": {
253
+ "@id": "benchmark-cases-parquet"
254
+ },
255
+ "extract": {
256
+ "column": "training_args.iit_weight"
257
+ }
258
+ }
259
+ },
260
+ {
261
+ "@type": "cr:Field",
262
+ "@id": "training_args.behavior_weight",
263
+ "name": "training_args.behavior_weight",
264
+ "description": "Column 'training_args.behavior_weight' from the parquet file describing all the cases in the benchmark.",
265
+ "dataType": "sc:Float",
266
+ "source": {
267
+ "fileSet": {
268
+ "@id": "benchmark-cases-parquet"
269
+ },
270
+ "extract": {
271
+ "column": "training_args.behavior_weight"
272
+ }
273
+ }
274
+ },
275
+ {
276
+ "@type": "cr:Field",
277
+ "@id": "training_args.strict_weight",
278
+ "name": "training_args.strict_weight",
279
+ "description": "Column 'training_args.strict_weight' from the parquet file describing all the cases in the benchmark.",
280
+ "dataType": "sc:Float",
281
+ "source": {
282
+ "fileSet": {
283
+ "@id": "benchmark-cases-parquet"
284
+ },
285
+ "extract": {
286
+ "column": "training_args.strict_weight"
287
+ }
288
+ }
289
+ },
290
+ {
291
+ "@type": "cr:Field",
292
+ "@id": "training_args.epochs",
293
+ "name": "training_args.epochs",
294
+ "description": "Column 'training_args.epochs' from the parquet file describing all the cases in the benchmark.",
295
+ "dataType": "sc:Float",
296
+ "source": {
297
+ "fileSet": {
298
+ "@id": "benchmark-cases-parquet"
299
+ },
300
+ "extract": {
301
+ "column": "training_args.epochs"
302
+ }
303
+ }
304
+ },
305
+ {
306
+ "@type": "cr:Field",
307
+ "@id": "training_args.act_fn",
308
+ "name": "training_args.act_fn",
309
+ "description": "Column 'training_args.act_fn' from the parquet file describing all the cases in the benchmark.",
310
+ "dataType": "sc:Text",
311
+ "source": {
312
+ "fileSet": {
313
+ "@id": "benchmark-cases-parquet"
314
+ },
315
+ "extract": {
316
+ "column": "training_args.act_fn"
317
+ }
318
+ }
319
+ },
320
+ {
321
+ "@type": "cr:Field",
322
+ "@id": "training_args.clip_grad_norm",
323
+ "name": "training_args.clip_grad_norm",
324
+ "description": "Column 'training_args.clip_grad_norm' from the parquet file describing all the cases in the benchmark.",
325
+ "dataType": "sc:Float",
326
+ "source": {
327
+ "fileSet": {
328
+ "@id": "benchmark-cases-parquet"
329
+ },
330
+ "extract": {
331
+ "column": "training_args.clip_grad_norm"
332
+ }
333
+ }
334
+ },
335
+ {
336
+ "@type": "cr:Field",
337
+ "@id": "training_args.lr_scheduler",
338
+ "name": "training_args.lr_scheduler",
339
+ "description": "Column 'training_args.lr_scheduler' from the parquet file describing all the cases in the benchmark.",
340
+ "dataType": "sc:Text",
341
+ "source": {
342
+ "fileSet": {
343
+ "@id": "benchmark-cases-parquet"
344
+ },
345
+ "extract": {
346
+ "column": "training_args.lr_scheduler"
347
+ }
348
+ }
349
+ },
350
+ {
351
+ "@type": "cr:Field",
352
+ "@id": "transformer_cfg.n_layers",
353
+ "name": "transformer_cfg.n_layers",
354
+ "description": "Column 'transformer_cfg.n_layers' from the parquet file describing all the cases in the benchmark.",
355
+ "dataType": "sc:Float",
356
+ "source": {
357
+ "fileSet": {
358
+ "@id": "benchmark-cases-parquet"
359
+ },
360
+ "extract": {
361
+ "column": "transformer_cfg.n_layers"
362
+ }
363
+ }
364
+ },
365
+ {
366
+ "@type": "cr:Field",
367
+ "@id": "transformer_cfg.d_model",
368
+ "name": "transformer_cfg.d_model",
369
+ "description": "Column 'transformer_cfg.d_model' from the parquet file describing all the cases in the benchmark.",
370
+ "dataType": "sc:Float",
371
+ "source": {
372
+ "fileSet": {
373
+ "@id": "benchmark-cases-parquet"
374
+ },
375
+ "extract": {
376
+ "column": "transformer_cfg.d_model"
377
+ }
378
+ }
379
+ },
380
+ {
381
+ "@type": "cr:Field",
382
+ "@id": "transformer_cfg.n_ctx",
383
+ "name": "transformer_cfg.n_ctx",
384
+ "description": "Column 'transformer_cfg.n_ctx' from the parquet file describing all the cases in the benchmark.",
385
+ "dataType": "sc:Float",
386
+ "source": {
387
+ "fileSet": {
388
+ "@id": "benchmark-cases-parquet"
389
+ },
390
+ "extract": {
391
+ "column": "transformer_cfg.n_ctx"
392
+ }
393
+ }
394
+ },
395
+ {
396
+ "@type": "cr:Field",
397
+ "@id": "transformer_cfg.d_head",
398
+ "name": "transformer_cfg.d_head",
399
+ "description": "Column 'transformer_cfg.d_head' from the parquet file describing all the cases in the benchmark.",
400
+ "dataType": "sc:Float",
401
+ "source": {
402
+ "fileSet": {
403
+ "@id": "benchmark-cases-parquet"
404
+ },
405
+ "extract": {
406
+ "column": "transformer_cfg.d_head"
407
+ }
408
+ }
409
+ },
410
+ {
411
+ "@type": "cr:Field",
412
+ "@id": "transformer_cfg.model_name",
413
+ "name": "transformer_cfg.model_name",
414
+ "description": "Column 'transformer_cfg.model_name' from the parquet file describing all the cases in the benchmark.",
415
+ "dataType": "sc:Text",
416
+ "source": {
417
+ "fileSet": {
418
+ "@id": "benchmark-cases-parquet"
419
+ },
420
+ "extract": {
421
+ "column": "transformer_cfg.model_name"
422
+ }
423
+ }
424
+ },
425
+ {
426
+ "@type": "cr:Field",
427
+ "@id": "transformer_cfg.n_heads",
428
+ "name": "transformer_cfg.n_heads",
429
+ "description": "Column 'transformer_cfg.n_heads' from the parquet file describing all the cases in the benchmark.",
430
+ "dataType": "sc:Float",
431
+ "source": {
432
+ "fileSet": {
433
+ "@id": "benchmark-cases-parquet"
434
+ },
435
+ "extract": {
436
+ "column": "transformer_cfg.n_heads"
437
+ }
438
+ }
439
+ },
440
+ {
441
+ "@type": "cr:Field",
442
+ "@id": "transformer_cfg.d_mlp",
443
+ "name": "transformer_cfg.d_mlp",
444
+ "description": "Column 'transformer_cfg.d_mlp' from the parquet file describing all the cases in the benchmark.",
445
+ "dataType": "sc:Float",
446
+ "source": {
447
+ "fileSet": {
448
+ "@id": "benchmark-cases-parquet"
449
+ },
450
+ "extract": {
451
+ "column": "transformer_cfg.d_mlp"
452
+ }
453
+ }
454
+ },
455
+ {
456
+ "@type": "cr:Field",
457
+ "@id": "transformer_cfg.act_fn",
458
+ "name": "transformer_cfg.act_fn",
459
+ "description": "Column 'transformer_cfg.act_fn' from the parquet file describing all the cases in the benchmark.",
460
+ "dataType": "sc:Text",
461
+ "source": {
462
+ "fileSet": {
463
+ "@id": "benchmark-cases-parquet"
464
+ },
465
+ "extract": {
466
+ "column": "transformer_cfg.act_fn"
467
+ }
468
+ }
469
+ },
470
+ {
471
+ "@type": "cr:Field",
472
+ "@id": "transformer_cfg.d_vocab",
473
+ "name": "transformer_cfg.d_vocab",
474
+ "description": "Column 'transformer_cfg.d_vocab' from the parquet file describing all the cases in the benchmark.",
475
+ "dataType": "sc:Float",
476
+ "source": {
477
+ "fileSet": {
478
+ "@id": "benchmark-cases-parquet"
479
+ },
480
+ "extract": {
481
+ "column": "transformer_cfg.d_vocab"
482
+ }
483
+ }
484
+ },
485
+ {
486
+ "@type": "cr:Field",
487
+ "@id": "transformer_cfg.eps",
488
+ "name": "transformer_cfg.eps",
489
+ "description": "Column 'transformer_cfg.eps' from the parquet file describing all the cases in the benchmark.",
490
+ "dataType": "sc:Float",
491
+ "source": {
492
+ "fileSet": {
493
+ "@id": "benchmark-cases-parquet"
494
+ },
495
+ "extract": {
496
+ "column": "transformer_cfg.eps"
497
+ }
498
+ }
499
+ },
500
+ {
501
+ "@type": "cr:Field",
502
+ "@id": "transformer_cfg.use_attn_result",
503
+ "name": "transformer_cfg.use_attn_result",
504
+ "description": "Column 'transformer_cfg.use_attn_result' from the parquet file describing all the cases in the benchmark.",
505
+ "dataType": "sc:Boolean",
506
+ "source": {
507
+ "fileSet": {
508
+ "@id": "benchmark-cases-parquet"
509
+ },
510
+ "extract": {
511
+ "column": "transformer_cfg.use_attn_result"
512
+ }
513
+ }
514
+ },
515
+ {
516
+ "@type": "cr:Field",
517
+ "@id": "transformer_cfg.use_attn_scale",
518
+ "name": "transformer_cfg.use_attn_scale",
519
+ "description": "Column 'transformer_cfg.use_attn_scale' from the parquet file describing all the cases in the benchmark.",
520
+ "dataType": "sc:Boolean",
521
+ "source": {
522
+ "fileSet": {
523
+ "@id": "benchmark-cases-parquet"
524
+ },
525
+ "extract": {
526
+ "column": "transformer_cfg.use_attn_scale"
527
+ }
528
+ }
529
+ },
530
+ {
531
+ "@type": "cr:Field",
532
+ "@id": "transformer_cfg.use_split_qkv_input",
533
+ "name": "transformer_cfg.use_split_qkv_input",
534
+ "description": "Column 'transformer_cfg.use_split_qkv_input' from the parquet file describing all the cases in the benchmark.",
535
+ "dataType": "sc:Boolean",
536
+ "source": {
537
+ "fileSet": {
538
+ "@id": "benchmark-cases-parquet"
539
+ },
540
+ "extract": {
541
+ "column": "transformer_cfg.use_split_qkv_input"
542
+ }
543
+ }
544
+ },
545
+ {
546
+ "@type": "cr:Field",
547
+ "@id": "transformer_cfg.use_hook_mlp_in",
548
+ "name": "transformer_cfg.use_hook_mlp_in",
549
+ "description": "Column 'transformer_cfg.use_hook_mlp_in' from the parquet file describing all the cases in the benchmark.",
550
+ "dataType": "sc:Boolean",
551
+ "source": {
552
+ "fileSet": {
553
+ "@id": "benchmark-cases-parquet"
554
+ },
555
+ "extract": {
556
+ "column": "transformer_cfg.use_hook_mlp_in"
557
+ }
558
+ }
559
+ },
560
+ {
561
+ "@type": "cr:Field",
562
+ "@id": "transformer_cfg.use_attn_in",
563
+ "name": "transformer_cfg.use_attn_in",
564
+ "description": "Column 'transformer_cfg.use_attn_in' from the parquet file describing all the cases in the benchmark.",
565
+ "dataType": "sc:Boolean",
566
+ "source": {
567
+ "fileSet": {
568
+ "@id": "benchmark-cases-parquet"
569
+ },
570
+ "extract": {
571
+ "column": "transformer_cfg.use_attn_in"
572
+ }
573
+ }
574
+ },
575
+ {
576
+ "@type": "cr:Field",
577
+ "@id": "transformer_cfg.use_local_attn",
578
+ "name": "transformer_cfg.use_local_attn",
579
+ "description": "Column 'transformer_cfg.use_local_attn' from the parquet file describing all the cases in the benchmark.",
580
+ "dataType": "sc:Boolean",
581
+ "source": {
582
+ "fileSet": {
583
+ "@id": "benchmark-cases-parquet"
584
+ },
585
+ "extract": {
586
+ "column": "transformer_cfg.use_local_attn"
587
+ }
588
+ }
589
+ },
590
+ {
591
+ "@type": "cr:Field",
592
+ "@id": "transformer_cfg.original_architecture",
593
+ "name": "transformer_cfg.original_architecture",
594
+ "description": "Column 'transformer_cfg.original_architecture' from the parquet file describing all the cases in the benchmark.",
595
+ "dataType": "sc:Float",
596
+ "source": {
597
+ "fileSet": {
598
+ "@id": "benchmark-cases-parquet"
599
+ },
600
+ "extract": {
601
+ "column": "transformer_cfg.original_architecture"
602
+ }
603
+ }
604
+ },
605
+ {
606
+ "@type": "cr:Field",
607
+ "@id": "transformer_cfg.from_checkpoint",
608
+ "name": "transformer_cfg.from_checkpoint",
609
+ "description": "Column 'transformer_cfg.from_checkpoint' from the parquet file describing all the cases in the benchmark.",
610
+ "dataType": "sc:Boolean",
611
+ "source": {
612
+ "fileSet": {
613
+ "@id": "benchmark-cases-parquet"
614
+ },
615
+ "extract": {
616
+ "column": "transformer_cfg.from_checkpoint"
617
+ }
618
+ }
619
+ },
620
+ {
621
+ "@type": "cr:Field",
622
+ "@id": "transformer_cfg.checkpoint_index",
623
+ "name": "transformer_cfg.checkpoint_index",
624
+ "description": "Column 'transformer_cfg.checkpoint_index' from the parquet file describing all the cases in the benchmark.",
625
+ "dataType": "sc:Float",
626
+ "source": {
627
+ "fileSet": {
628
+ "@id": "benchmark-cases-parquet"
629
+ },
630
+ "extract": {
631
+ "column": "transformer_cfg.checkpoint_index"
632
+ }
633
+ }
634
+ },
635
+ {
636
+ "@type": "cr:Field",
637
+ "@id": "transformer_cfg.checkpoint_label_type",
638
+ "name": "transformer_cfg.checkpoint_label_type",
639
+ "description": "Column 'transformer_cfg.checkpoint_label_type' from the parquet file describing all the cases in the benchmark.",
640
+ "dataType": "sc:Float",
641
+ "source": {
642
+ "fileSet": {
643
+ "@id": "benchmark-cases-parquet"
644
+ },
645
+ "extract": {
646
+ "column": "transformer_cfg.checkpoint_label_type"
647
+ }
648
+ }
649
+ },
650
+ {
651
+ "@type": "cr:Field",
652
+ "@id": "transformer_cfg.checkpoint_value",
653
+ "name": "transformer_cfg.checkpoint_value",
654
+ "description": "Column 'transformer_cfg.checkpoint_value' from the parquet file describing all the cases in the benchmark.",
655
+ "dataType": "sc:Float",
656
+ "source": {
657
+ "fileSet": {
658
+ "@id": "benchmark-cases-parquet"
659
+ },
660
+ "extract": {
661
+ "column": "transformer_cfg.checkpoint_value"
662
+ }
663
+ }
664
+ },
665
+ {
666
+ "@type": "cr:Field",
667
+ "@id": "transformer_cfg.tokenizer_name",
668
+ "name": "transformer_cfg.tokenizer_name",
669
+ "description": "Column 'transformer_cfg.tokenizer_name' from the parquet file describing all the cases in the benchmark.",
670
+ "dataType": "sc:Float",
671
+ "source": {
672
+ "fileSet": {
673
+ "@id": "benchmark-cases-parquet"
674
+ },
675
+ "extract": {
676
+ "column": "transformer_cfg.tokenizer_name"
677
+ }
678
+ }
679
+ },
680
+ {
681
+ "@type": "cr:Field",
682
+ "@id": "transformer_cfg.window_size",
683
+ "name": "transformer_cfg.window_size",
684
+ "description": "Column 'transformer_cfg.window_size' from the parquet file describing all the cases in the benchmark.",
685
+ "dataType": "sc:Float",
686
+ "source": {
687
+ "fileSet": {
688
+ "@id": "benchmark-cases-parquet"
689
+ },
690
+ "extract": {
691
+ "column": "transformer_cfg.window_size"
692
+ }
693
+ }
694
+ },
695
+ {
696
+ "@type": "cr:Field",
697
+ "@id": "transformer_cfg.attn_types",
698
+ "name": "transformer_cfg.attn_types",
699
+ "description": "Column 'transformer_cfg.attn_types' from the parquet file describing all the cases in the benchmark.",
700
+ "dataType": "sc:Float",
701
+ "source": {
702
+ "fileSet": {
703
+ "@id": "benchmark-cases-parquet"
704
+ },
705
+ "extract": {
706
+ "column": "transformer_cfg.attn_types"
707
+ }
708
+ }
709
+ },
710
+ {
711
+ "@type": "cr:Field",
712
+ "@id": "transformer_cfg.init_mode",
713
+ "name": "transformer_cfg.init_mode",
714
+ "description": "Column 'transformer_cfg.init_mode' from the parquet file describing all the cases in the benchmark.",
715
+ "dataType": "sc:Text",
716
+ "source": {
717
+ "fileSet": {
718
+ "@id": "benchmark-cases-parquet"
719
+ },
720
+ "extract": {
721
+ "column": "transformer_cfg.init_mode"
722
+ }
723
+ }
724
+ },
725
+ {
726
+ "@type": "cr:Field",
727
+ "@id": "transformer_cfg.normalization_type",
728
+ "name": "transformer_cfg.normalization_type",
729
+ "description": "Column 'transformer_cfg.normalization_type' from the parquet file describing all the cases in the benchmark.",
730
+ "dataType": "sc:Float",
731
+ "source": {
732
+ "fileSet": {
733
+ "@id": "benchmark-cases-parquet"
734
+ },
735
+ "extract": {
736
+ "column": "transformer_cfg.normalization_type"
737
+ }
738
+ }
739
+ },
740
+ {
741
+ "@type": "cr:Field",
742
+ "@id": "transformer_cfg.device",
743
+ "name": "transformer_cfg.device",
744
+ "description": "Column 'transformer_cfg.device' from the parquet file describing all the cases in the benchmark.",
745
+ "dataType": "sc:Text",
746
+ "source": {
747
+ "fileSet": {
748
+ "@id": "benchmark-cases-parquet"
749
+ },
750
+ "extract": {
751
+ "column": "transformer_cfg.device"
752
+ }
753
+ }
754
+ },
755
+ {
756
+ "@type": "cr:Field",
757
+ "@id": "transformer_cfg.n_devices",
758
+ "name": "transformer_cfg.n_devices",
759
+ "description": "Column 'transformer_cfg.n_devices' from the parquet file describing all the cases in the benchmark.",
760
+ "dataType": "sc:Float",
761
+ "source": {
762
+ "fileSet": {
763
+ "@id": "benchmark-cases-parquet"
764
+ },
765
+ "extract": {
766
+ "column": "transformer_cfg.n_devices"
767
+ }
768
+ }
769
+ },
770
+ {
771
+ "@type": "cr:Field",
772
+ "@id": "transformer_cfg.attention_dir",
773
+ "name": "transformer_cfg.attention_dir",
774
+ "description": "Column 'transformer_cfg.attention_dir' from the parquet file describing all the cases in the benchmark.",
775
+ "dataType": "sc:Text",
776
+ "source": {
777
+ "fileSet": {
778
+ "@id": "benchmark-cases-parquet"
779
+ },
780
+ "extract": {
781
+ "column": "transformer_cfg.attention_dir"
782
+ }
783
+ }
784
+ },
785
+ {
786
+ "@type": "cr:Field",
787
+ "@id": "transformer_cfg.attn_only",
788
+ "name": "transformer_cfg.attn_only",
789
+ "description": "Column 'transformer_cfg.attn_only' from the parquet file describing all the cases in the benchmark.",
790
+ "dataType": "sc:Boolean",
791
+ "source": {
792
+ "fileSet": {
793
+ "@id": "benchmark-cases-parquet"
794
+ },
795
+ "extract": {
796
+ "column": "transformer_cfg.attn_only"
797
+ }
798
+ }
799
+ },
800
+ {
801
+ "@type": "cr:Field",
802
+ "@id": "transformer_cfg.seed",
803
+ "name": "transformer_cfg.seed",
804
+ "description": "Column 'transformer_cfg.seed' from the parquet file describing all the cases in the benchmark.",
805
+ "dataType": "sc:Float",
806
+ "source": {
807
+ "fileSet": {
808
+ "@id": "benchmark-cases-parquet"
809
+ },
810
+ "extract": {
811
+ "column": "transformer_cfg.seed"
812
+ }
813
+ }
814
+ },
815
+ {
816
+ "@type": "cr:Field",
817
+ "@id": "transformer_cfg.initializer_range",
818
+ "name": "transformer_cfg.initializer_range",
819
+ "description": "Column 'transformer_cfg.initializer_range' from the parquet file describing all the cases in the benchmark.",
820
+ "dataType": "sc:Float",
821
+ "source": {
822
+ "fileSet": {
823
+ "@id": "benchmark-cases-parquet"
824
+ },
825
+ "extract": {
826
+ "column": "transformer_cfg.initializer_range"
827
+ }
828
+ }
829
+ },
830
+ {
831
+ "@type": "cr:Field",
832
+ "@id": "transformer_cfg.init_weights",
833
+ "name": "transformer_cfg.init_weights",
834
+ "description": "Column 'transformer_cfg.init_weights' from the parquet file describing all the cases in the benchmark.",
835
+ "dataType": "sc:Boolean",
836
+ "source": {
837
+ "fileSet": {
838
+ "@id": "benchmark-cases-parquet"
839
+ },
840
+ "extract": {
841
+ "column": "transformer_cfg.init_weights"
842
+ }
843
+ }
844
+ },
845
+ {
846
+ "@type": "cr:Field",
847
+ "@id": "transformer_cfg.scale_attn_by_inverse_layer_idx",
848
+ "name": "transformer_cfg.scale_attn_by_inverse_layer_idx",
849
+ "description": "Column 'transformer_cfg.scale_attn_by_inverse_layer_idx' from the parquet file describing all the cases in the benchmark.",
850
+ "dataType": "sc:Boolean",
851
+ "source": {
852
+ "fileSet": {
853
+ "@id": "benchmark-cases-parquet"
854
+ },
855
+ "extract": {
856
+ "column": "transformer_cfg.scale_attn_by_inverse_layer_idx"
857
+ }
858
+ }
859
+ },
860
+ {
861
+ "@type": "cr:Field",
862
+ "@id": "transformer_cfg.positional_embedding_type",
863
+ "name": "transformer_cfg.positional_embedding_type",
864
+ "description": "Column 'transformer_cfg.positional_embedding_type' from the parquet file describing all the cases in the benchmark.",
865
+ "dataType": "sc:Text",
866
+ "source": {
867
+ "fileSet": {
868
+ "@id": "benchmark-cases-parquet"
869
+ },
870
+ "extract": {
871
+ "column": "transformer_cfg.positional_embedding_type"
872
+ }
873
+ }
874
+ },
875
+ {
876
+ "@type": "cr:Field",
877
+ "@id": "transformer_cfg.final_rms",
878
+ "name": "transformer_cfg.final_rms",
879
+ "description": "Column 'transformer_cfg.final_rms' from the parquet file describing all the cases in the benchmark.",
880
+ "dataType": "sc:Boolean",
881
+ "source": {
882
+ "fileSet": {
883
+ "@id": "benchmark-cases-parquet"
884
+ },
885
+ "extract": {
886
+ "column": "transformer_cfg.final_rms"
887
+ }
888
+ }
889
+ },
890
+ {
891
+ "@type": "cr:Field",
892
+ "@id": "transformer_cfg.d_vocab_out",
893
+ "name": "transformer_cfg.d_vocab_out",
894
+ "description": "Column 'transformer_cfg.d_vocab_out' from the parquet file describing all the cases in the benchmark.",
895
+ "dataType": "sc:Float",
896
+ "source": {
897
+ "fileSet": {
898
+ "@id": "benchmark-cases-parquet"
899
+ },
900
+ "extract": {
901
+ "column": "transformer_cfg.d_vocab_out"
902
+ }
903
+ }
904
+ },
905
+ {
906
+ "@type": "cr:Field",
907
+ "@id": "transformer_cfg.parallel_attn_mlp",
908
+ "name": "transformer_cfg.parallel_attn_mlp",
909
+ "description": "Column 'transformer_cfg.parallel_attn_mlp' from the parquet file describing all the cases in the benchmark.",
910
+ "dataType": "sc:Boolean",
911
+ "source": {
912
+ "fileSet": {
913
+ "@id": "benchmark-cases-parquet"
914
+ },
915
+ "extract": {
916
+ "column": "transformer_cfg.parallel_attn_mlp"
917
+ }
918
+ }
919
+ },
920
+ {
921
+ "@type": "cr:Field",
922
+ "@id": "transformer_cfg.rotary_dim",
923
+ "name": "transformer_cfg.rotary_dim",
924
+ "description": "Column 'transformer_cfg.rotary_dim' from the parquet file describing all the cases in the benchmark.",
925
+ "dataType": "sc:Float",
926
+ "source": {
927
+ "fileSet": {
928
+ "@id": "benchmark-cases-parquet"
929
+ },
930
+ "extract": {
931
+ "column": "transformer_cfg.rotary_dim"
932
+ }
933
+ }
934
+ },
935
+ {
936
+ "@type": "cr:Field",
937
+ "@id": "transformer_cfg.n_params",
938
+ "name": "transformer_cfg.n_params",
939
+ "description": "Column 'transformer_cfg.n_params' from the parquet file describing all the cases in the benchmark.",
940
+ "dataType": "sc:Float",
941
+ "source": {
942
+ "fileSet": {
943
+ "@id": "benchmark-cases-parquet"
944
+ },
945
+ "extract": {
946
+ "column": "transformer_cfg.n_params"
947
+ }
948
+ }
949
+ },
950
+ {
951
+ "@type": "cr:Field",
952
+ "@id": "transformer_cfg.use_hook_tokens",
953
+ "name": "transformer_cfg.use_hook_tokens",
954
+ "description": "Column 'transformer_cfg.use_hook_tokens' from the parquet file describing all the cases in the benchmark.",
955
+ "dataType": "sc:Boolean",
956
+ "source": {
957
+ "fileSet": {
958
+ "@id": "benchmark-cases-parquet"
959
+ },
960
+ "extract": {
961
+ "column": "transformer_cfg.use_hook_tokens"
962
+ }
963
+ }
964
+ },
965
+ {
966
+ "@type": "cr:Field",
967
+ "@id": "transformer_cfg.gated_mlp",
968
+ "name": "transformer_cfg.gated_mlp",
969
+ "description": "Column 'transformer_cfg.gated_mlp' from the parquet file describing all the cases in the benchmark.",
970
+ "dataType": "sc:Boolean",
971
+ "source": {
972
+ "fileSet": {
973
+ "@id": "benchmark-cases-parquet"
974
+ },
975
+ "extract": {
976
+ "column": "transformer_cfg.gated_mlp"
977
+ }
978
+ }
979
+ },
980
+ {
981
+ "@type": "cr:Field",
982
+ "@id": "transformer_cfg.default_prepend_bos",
983
+ "name": "transformer_cfg.default_prepend_bos",
984
+ "description": "Column 'transformer_cfg.default_prepend_bos' from the parquet file describing all the cases in the benchmark.",
985
+ "dataType": "sc:Boolean",
986
+ "source": {
987
+ "fileSet": {
988
+ "@id": "benchmark-cases-parquet"
989
+ },
990
+ "extract": {
991
+ "column": "transformer_cfg.default_prepend_bos"
992
+ }
993
+ }
994
+ },
995
+ {
996
+ "@type": "cr:Field",
997
+ "@id": "transformer_cfg.dtype",
998
+ "name": "transformer_cfg.dtype",
999
+ "description": "Column 'transformer_cfg.dtype' from the parquet file describing all the cases in the benchmark.",
1000
+ "dataType": "sc:Text",
1001
+ "source": {
1002
+ "fileSet": {
1003
+ "@id": "benchmark-cases-parquet"
1004
+ },
1005
+ "extract": {
1006
+ "column": "transformer_cfg.dtype"
1007
+ }
1008
+ }
1009
+ },
1010
+ {
1011
+ "@type": "cr:Field",
1012
+ "@id": "transformer_cfg.tokenizer_prepends_bos",
1013
+ "name": "transformer_cfg.tokenizer_prepends_bos",
1014
+ "description": "Column 'transformer_cfg.tokenizer_prepends_bos' from the parquet file describing all the cases in the benchmark.",
1015
+ "dataType": "sc:Float",
1016
+ "source": {
1017
+ "fileSet": {
1018
+ "@id": "benchmark-cases-parquet"
1019
+ },
1020
+ "extract": {
1021
+ "column": "transformer_cfg.tokenizer_prepends_bos"
1022
+ }
1023
+ }
1024
+ },
1025
+ {
1026
+ "@type": "cr:Field",
1027
+ "@id": "transformer_cfg.n_key_value_heads",
1028
+ "name": "transformer_cfg.n_key_value_heads",
1029
+ "description": "Column 'transformer_cfg.n_key_value_heads' from the parquet file describing all the cases in the benchmark.",
1030
+ "dataType": "sc:Float",
1031
+ "source": {
1032
+ "fileSet": {
1033
+ "@id": "benchmark-cases-parquet"
1034
+ },
1035
+ "extract": {
1036
+ "column": "transformer_cfg.n_key_value_heads"
1037
+ }
1038
+ }
1039
+ },
1040
+ {
1041
+ "@type": "cr:Field",
1042
+ "@id": "transformer_cfg.post_embedding_ln",
1043
+ "name": "transformer_cfg.post_embedding_ln",
1044
+ "description": "Column 'transformer_cfg.post_embedding_ln' from the parquet file describing all the cases in the benchmark.",
1045
+ "dataType": "sc:Boolean",
1046
+ "source": {
1047
+ "fileSet": {
1048
+ "@id": "benchmark-cases-parquet"
1049
+ },
1050
+ "extract": {
1051
+ "column": "transformer_cfg.post_embedding_ln"
1052
+ }
1053
+ }
1054
+ },
1055
+ {
1056
+ "@type": "cr:Field",
1057
+ "@id": "transformer_cfg.rotary_base",
1058
+ "name": "transformer_cfg.rotary_base",
1059
+ "description": "Column 'transformer_cfg.rotary_base' from the parquet file describing all the cases in the benchmark.",
1060
+ "dataType": "sc:Float",
1061
+ "source": {
1062
+ "fileSet": {
1063
+ "@id": "benchmark-cases-parquet"
1064
+ },
1065
+ "extract": {
1066
+ "column": "transformer_cfg.rotary_base"
1067
+ }
1068
+ }
1069
+ },
1070
+ {
1071
+ "@type": "cr:Field",
1072
+ "@id": "transformer_cfg.trust_remote_code",
1073
+ "name": "transformer_cfg.trust_remote_code",
1074
+ "description": "Column 'transformer_cfg.trust_remote_code' from the parquet file describing all the cases in the benchmark.",
1075
+ "dataType": "sc:Boolean",
1076
+ "source": {
1077
+ "fileSet": {
1078
+ "@id": "benchmark-cases-parquet"
1079
+ },
1080
+ "extract": {
1081
+ "column": "transformer_cfg.trust_remote_code"
1082
+ }
1083
+ }
1084
+ },
1085
+ {
1086
+ "@type": "cr:Field",
1087
+ "@id": "transformer_cfg.rotary_adjacent_pairs",
1088
+ "name": "transformer_cfg.rotary_adjacent_pairs",
1089
+ "description": "Column 'transformer_cfg.rotary_adjacent_pairs' from the parquet file describing all the cases in the benchmark.",
1090
+ "dataType": "sc:Boolean",
1091
+ "source": {
1092
+ "fileSet": {
1093
+ "@id": "benchmark-cases-parquet"
1094
+ },
1095
+ "extract": {
1096
+ "column": "transformer_cfg.rotary_adjacent_pairs"
1097
+ }
1098
+ }
1099
+ }
1100
+ ]
1101
+ }
1102
+ ]
1103
+ }