ilu000 commited on
Commit
ce0ab8f
1 Parent(s): 47b854a

Upload 9 files

Browse files
README.md ADDED
@@ -0,0 +1,145 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ language:
3
+ - en
4
+ library_name: transformers
5
+ license: apache-2.0
6
+ tags:
7
+ - gpt
8
+ - llm
9
+ - large language model
10
+ - h2o-llmstudio
11
+ inference: false
12
+ thumbnail: https://h2o.ai/etc.clientlibs/h2o/clientlibs/clientlib-site/resources/images/favicon.ico
13
+ ---
14
+ # Model Card
15
+ ## Summary
16
+
17
+ h2o-danube-1.8b-chat is an chat fine-tuned model by H2O.ai with 1.8 billion parameters. We release three versions of this model:
18
+
19
+ | Model Name | Description |
20
+ |:-----------------------------------------------------------------------------------|:----------------|
21
+ | [h2oai/h2o-danube-1.8b-base](https://huggingface.co/h2oai/h2o-danube-1.8b-base) | Base model |
22
+ | [h2oai/h2o-danube-1.8b-sft](https://huggingface.co/h2oai/h2o-danube-1.8b-sft) | SFT tuned |
23
+ | [h2oai/h2o-danube-1.8b-chat](https://huggingface.co/h2oai/h2o-danube-1.8b-chat) | SFT + DPO tuned |
24
+
25
+ This model was trained using [H2O LLM Studio](https://github.com/h2oai/h2o-llmstudio).
26
+
27
+ ## Model Architecture
28
+
29
+ We adjust the Llama 2 architecture for a total of around 1.8b parameters. We use the original Llama 2 tokenizer with a vocabulary size of 32,000 and train our model up to a context length of 16,384. We incorporate the sliding window attention from mistral with a size of 4,096.
30
+
31
+ The details of the model architecture are:
32
+
33
+ | Hyperparameter | Value |
34
+ |:----------------|:-------|
35
+ | n_layers | 24 |
36
+ | n_heads | 32 |
37
+ | n_query_groups | 8 |
38
+ | n_embd | 2560 |
39
+ | vocab size | 32000 |
40
+ | sequence length | 16384 |
41
+
42
+ ## Usage
43
+
44
+ To use the model with the `transformers` library on a machine with GPUs, first make sure you have the `transformers` library installed.
45
+
46
+ ```bash
47
+ pip install transformers==4.36.1
48
+ ```
49
+
50
+ ```python
51
+ import torch
52
+ from transformers import pipeline
53
+
54
+ pipe = pipeline(
55
+ "text-generation",
56
+ model="h2oai/h2o-danube-1.8b-chat",
57
+ torch_dtype=torch.bfloat16,
58
+ device_map="auto",
59
+ )
60
+
61
+ # We use the HF Tokenizer chat template to format each message
62
+ # https://huggingface.co/docs/transformers/main/en/chat_templating
63
+ messages = [
64
+ {"role": "user", "content": "Why is drinking water so healthy?"},
65
+ ]
66
+ prompt = pipe.tokenizer.apply_chat_template(
67
+ messages,
68
+ tokenize=False,
69
+ add_generation_prompt=True,
70
+ )
71
+ res = pipe(
72
+ prompt,
73
+ max_new_tokens=256,
74
+ )
75
+ print(res[0]["generated_text"])
76
+ # <|system|>You are a friendly chatbot</s><|prompt|>Why is drinking water so healthy?</s><|answer|> Drinking water is healthy for several reasons: [...]
77
+ ```
78
+
79
+ ## Benchmarks
80
+
81
+ Commonsense, world-knowledge and reading comprehension tested in 0-shot:
82
+
83
+ | Benchmark | acc_n |
84
+ |:--------------|:--------:|
85
+ | ARC-easy | 67.51 |
86
+ | ARC-challenge | 39.25 |
87
+ | BoolQ | 77.89 |
88
+ | Hellaswag | 67.60 |
89
+ | OpenBookQA | 39.20 |
90
+ | PiQA | 76.71 |
91
+ | TriviaQA | 36.29 |
92
+ | Winogrande | 65.35 |
93
+
94
+ ## Quantization and sharding
95
+
96
+ You can load the models using quantization by specifying ```load_in_8bit=True``` or ```load_in_4bit=True```. Also, sharding on multiple GPUs is possible by setting ```device_map=auto```.
97
+
98
+ ## Model Architecture
99
+
100
+ ```
101
+ MistralForCausalLM(
102
+ (model): MistralModel(
103
+ (embed_tokens): Embedding(32000, 2560, padding_idx=0)
104
+ (layers): ModuleList(
105
+ (0-23): 24 x MistralDecoderLayer(
106
+ (self_attn): MistralAttention(
107
+ (q_proj): Linear(in_features=2560, out_features=2560, bias=False)
108
+ (k_proj): Linear(in_features=2560, out_features=640, bias=False)
109
+ (v_proj): Linear(in_features=2560, out_features=640, bias=False)
110
+ (o_proj): Linear(in_features=2560, out_features=2560, bias=False)
111
+ (rotary_emb): MistralRotaryEmbedding()
112
+ )
113
+ (mlp): MistralMLP(
114
+ (gate_proj): Linear(in_features=2560, out_features=6912, bias=False)
115
+ (up_proj): Linear(in_features=2560, out_features=6912, bias=False)
116
+ (down_proj): Linear(in_features=6912, out_features=2560, bias=False)
117
+ (act_fn): SiLU()
118
+ )
119
+ (input_layernorm): MistralRMSNorm()
120
+ (post_attention_layernorm): MistralRMSNorm()
121
+ )
122
+ )
123
+ (norm): MistralRMSNorm()
124
+ )
125
+ (lm_head): Linear(in_features=2560, out_features=32000, bias=False)
126
+ )
127
+ ```
128
+
129
+ ## Model Configuration
130
+
131
+ This model was trained using H2O LLM Studio and with the configuration in [cfg.yaml](cfg.yaml). Visit [H2O LLM Studio](https://github.com/h2oai/h2o-llmstudio) to learn how to train your own large language models.
132
+
133
+
134
+ ## Disclaimer
135
+
136
+ Please read this disclaimer carefully before using the large language model provided in this repository. Your use of the model signifies your agreement to the following terms and conditions.
137
+
138
+ - Biases and Offensiveness: The large language model is trained on a diverse range of internet text data, which may contain biased, racist, offensive, or otherwise inappropriate content. By using this model, you acknowledge and accept that the generated content may sometimes exhibit biases or produce content that is offensive or inappropriate. The developers of this repository do not endorse, support, or promote any such content or viewpoints.
139
+ - Limitations: The large language model is an AI-based tool and not a human. It may produce incorrect, nonsensical, or irrelevant responses. It is the user's responsibility to critically evaluate the generated content and use it at their discretion.
140
+ - Use at Your Own Risk: Users of this large language model must assume full responsibility for any consequences that may arise from their use of the tool. The developers and contributors of this repository shall not be held liable for any damages, losses, or harm resulting from the use or misuse of the provided model.
141
+ - Ethical Considerations: Users are encouraged to use the large language model responsibly and ethically. By using this model, you agree not to use it for purposes that promote hate speech, discrimination, harassment, or any form of illegal or harmful activities.
142
+ - Reporting Issues: If you encounter any biased, offensive, or otherwise inappropriate content generated by the large language model, please report it to the repository maintainers through the provided channels. Your feedback will help improve the model and mitigate potential issues.
143
+ - Changes to this Disclaimer: The developers of this repository reserve the right to modify or update this disclaimer at any time without prior notice. It is the user's responsibility to periodically review the disclaimer to stay informed about any changes.
144
+
145
+ By using the large language model provided in this repository, you agree to accept and comply with the terms and conditions outlined in this disclaimer. If you do not agree with any part of this disclaimer, you should refrain from using the model and any content generated by it.
cfg.yaml ADDED
@@ -0,0 +1,109 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ architecture:
2
+ backbone_dtype: bfloat16
3
+ force_embedding_gradients: false
4
+ gradient_checkpointing: true
5
+ intermediate_dropout: 0.0
6
+ pretrained: true
7
+ pretrained_weights: ''
8
+ augmentation:
9
+ neftune_noise_alpha: 0.0
10
+ random_parent_probability: 0.0
11
+ skip_parent_probability: 0.0
12
+ token_mask_probability: 0.0
13
+ dataset:
14
+ add_eos_token_to_answer: true
15
+ add_eos_token_to_prompt: true
16
+ add_eos_token_to_system: true
17
+ answer_column: output_min_rank
18
+ chatbot_author: H2O.ai
19
+ chatbot_name: h2oGPT
20
+ data_sample: 1.0
21
+ data_sample_choice:
22
+ - Train
23
+ - Validation
24
+ limit_chained_samples: true
25
+ mask_prompt_labels: true
26
+ parent_id_column: parent_id
27
+ personalize: false
28
+ prompt_column:
29
+ - instruction
30
+ rejected_answer_column: output_max_rank
31
+ system_column: None
32
+ text_answer_separator: <|answer|>
33
+ text_prompt_start: <|prompt|>
34
+ text_system_start: <|system|>
35
+ train_dataframe: dpo.pq
36
+ validation_dataframe: None
37
+ validation_size: 0.01
38
+ validation_strategy: automatic
39
+ environment:
40
+ compile_model: false
41
+ deepspeed_reduce_bucket_size: 1000000
42
+ deepspeed_stage3_param_persistence_threshold: 1000000
43
+ deepspeed_stage3_prefetch_bucket_size: 1000000
44
+ find_unused_parameters: false
45
+ gpus:
46
+ - '0'
47
+ - '1'
48
+ - '2'
49
+ huggingface_branch: main
50
+ mixed_precision: true
51
+ number_of_workers: 8
52
+ seed: -1
53
+ trust_remote_code: true
54
+ use_deepspeed: false
55
+ experiment_name: h2o-danube-1.8b-chat
56
+ llm_backbone: h2oai/h2o-danube-1.8b-base
57
+ logging:
58
+ logger: Neptune
59
+ neptune_project: ''
60
+ output_directory: output/
61
+ prediction:
62
+ batch_size_inference: 0
63
+ do_sample: false
64
+ max_length_inference: 256
65
+ metric: Perplexity
66
+ metric_gpt_model: gpt-3.5-turbo-0301
67
+ metric_gpt_template: general
68
+ min_length_inference: 2
69
+ num_beams: 1
70
+ num_history: 4
71
+ repetition_penalty: 1.0
72
+ stop_tokens: ''
73
+ temperature: 0.0
74
+ top_k: 0
75
+ top_p: 1.0
76
+ problem_type: text_dpo_modeling
77
+ tokenizer:
78
+ add_prefix_space: false
79
+ add_prompt_answer_tokens: false
80
+ max_length: 16384
81
+ max_length_answer: 8192
82
+ max_length_prompt: 8192
83
+ padding_quantile: 1.0
84
+ use_fast: true
85
+ training:
86
+ batch_size: 2
87
+ beta: 0.2
88
+ differential_learning_rate: 1.0e-05
89
+ differential_learning_rate_layers: []
90
+ drop_last_batch: true
91
+ epochs: 1
92
+ evaluate_before_training: false
93
+ evaluation_epochs: 1.0
94
+ grad_accumulation: 1
95
+ gradient_clip: 10.0
96
+ learning_rate: 3.0e-06
97
+ lora: true
98
+ lora_alpha: 16
99
+ lora_dropout: 0.05
100
+ lora_r: 4
101
+ lora_target_modules: ''
102
+ loss_function: DPOLoss
103
+ optimizer: AdamW
104
+ save_best_checkpoint: false
105
+ schedule: Cosine
106
+ train_validation_data: false
107
+ use_flash_attention_2: true
108
+ warmup_epochs: 0.05
109
+ weight_decay: 0.0
config.json ADDED
@@ -0,0 +1,29 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "h2oai/h2o-danube-1.8b-chat",
3
+ "architectures": [
4
+ "MistralForCausalLM"
5
+ ],
6
+ "attention_dropout": 0.0,
7
+ "bos_token_id": 1,
8
+ "eos_token_id": 2,
9
+ "hidden_act": "silu",
10
+ "hidden_size": 2560,
11
+ "initializer_range": 0.02,
12
+ "intermediate_size": 6912,
13
+ "max_position_embeddings": 16384,
14
+ "model_type": "mistral",
15
+ "num_attention_heads": 32,
16
+ "num_hidden_layers": 24,
17
+ "num_key_value_heads": 8,
18
+ "pad_token_id": 0,
19
+ "pretraining_tp": 1,
20
+ "rms_norm_eps": 1e-05,
21
+ "rope_scaling": null,
22
+ "rope_theta": 10000.0,
23
+ "sliding_window": 4096,
24
+ "tie_word_embeddings": false,
25
+ "torch_dtype": "bfloat16",
26
+ "transformers_version": "4.36.2",
27
+ "use_cache": true,
28
+ "vocab_size": 32000
29
+ }
generation_config.json ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_from_model_config": true,
3
+ "bos_token_id": 1,
4
+ "eos_token_id": 2,
5
+ "pad_token_id": 0,
6
+ "transformers_version": "4.36.2",
7
+ "repetition_penalty": 1.1
8
+ }
model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2a9a72b0fb854d00c0f624f640376ba95c46746350187c6352bf5f4159c8d936
3
+ size 3662427808
special_tokens_map.json ADDED
@@ -0,0 +1,44 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "bos_token": {
3
+ "content": "<s>",
4
+ "lstrip": false,
5
+ "normalized": false,
6
+ "rstrip": false,
7
+ "single_word": false
8
+ },
9
+ "cls_token": {
10
+ "content": "</s>",
11
+ "lstrip": false,
12
+ "normalized": false,
13
+ "rstrip": false,
14
+ "single_word": false
15
+ },
16
+ "eos_token": {
17
+ "content": "</s>",
18
+ "lstrip": false,
19
+ "normalized": false,
20
+ "rstrip": false,
21
+ "single_word": false
22
+ },
23
+ "pad_token": {
24
+ "content": "<unk>",
25
+ "lstrip": false,
26
+ "normalized": false,
27
+ "rstrip": false,
28
+ "single_word": false
29
+ },
30
+ "sep_token": {
31
+ "content": "</s>",
32
+ "lstrip": false,
33
+ "normalized": false,
34
+ "rstrip": false,
35
+ "single_word": false
36
+ },
37
+ "unk_token": {
38
+ "content": "<unk>",
39
+ "lstrip": false,
40
+ "normalized": false,
41
+ "rstrip": false,
42
+ "single_word": false
43
+ }
44
+ }
tokenizer.json ADDED
The diff for this file is too large to render. See raw diff
 
tokenizer.model ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9e556afd44213b6bd1be2b850ebbbd98f5481437a8021afaf58ee7fb1818d347
3
+ size 499723
tokenizer_config.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_bos_token": true,
3
+ "add_eos_token": false,
4
+ "add_prefix_space": false,
5
+ "added_tokens_decoder": {
6
+ "0": {
7
+ "content": "<unk>",
8
+ "lstrip": false,
9
+ "normalized": false,
10
+ "rstrip": false,
11
+ "single_word": false,
12
+ "special": true
13
+ },
14
+ "1": {
15
+ "content": "<s>",
16
+ "lstrip": false,
17
+ "normalized": false,
18
+ "rstrip": false,
19
+ "single_word": false,
20
+ "special": true
21
+ },
22
+ "2": {
23
+ "content": "</s>",
24
+ "lstrip": false,
25
+ "normalized": false,
26
+ "rstrip": false,
27
+ "single_word": false,
28
+ "special": true
29
+ }
30
+ },
31
+ "bos_token": "<s>",
32
+ "chat_template": "{% for message in messages %}{% if message['role'] == 'user' %}{{ '<|prompt|>' + message['content'] + eos_token }}{% elif message['role'] == 'system' %}{{ '<|system|>' + message['content'] + eos_token }}{% elif message['role'] == 'assistant' %}{{ '<|answer|>' + message['content'] + eos_token }}{% endif %}{% if loop.last and add_generation_prompt %}{{ '<|answer|>' }}{% endif %}{% endfor %}",
33
+ "clean_up_tokenization_spaces": false,
34
+ "cls_token": "</s>",
35
+ "eos_token": "</s>",
36
+ "legacy": false,
37
+ "model_max_length": 1000000000000000019884624838656,
38
+ "pad_token": "<unk>",
39
+ "padding_side": "left",
40
+ "sep_token": "</s>",
41
+ "sp_model_kwargs": {},
42
+ "tokenizer_class": "LlamaTokenizer",
43
+ "unk_token": "<unk>",
44
+ "use_default_system_prompt": false
45
+ }