Update README.md

README.md

@@ -17,6 +17,173 @@ tags:
 This model was converted to GGUF format from [`allura-org/TQ2.5-14B-Sugarquill-v1`](https://huggingface.co/allura-org/TQ2.5-14B-Sugarquill-v1) using llama.cpp via the ggml.ai's [GGUF-my-repo](https://huggingface.co/spaces/ggml-org/gguf-my-repo) space.
 Refer to the [original model card](https://huggingface.co/allura-org/TQ2.5-14B-Sugarquill-v1) for more details on the model.
 
+---
+Model details:
+-
+Qwen2.5-14B Sugarquill v1
+
+A continued pretrain of SuperNova-Medius on assorted short story data from the web. Supernova already had a nice prose, but diversifying it a bit definitely doesn't hurt. Also, finally a storywriter model with enough context for something more than a short story, that's also nice.
+
+It's a fair bit more temperamental than Gemma, but can be tamed with some sampling. Instruction following also stayed rather strong, so it works for both RP and storywriting, both in chat mode via back-and-forth co-writing and on raw completion.
+
+Overall, I'd say it successfully transfers the essence of what I liked about Gemma Sugarquill. I will also make a Qwen version of Aletheia, but with a brand new LoRA, based on a brand new RP dataset that's in the making right now.
+
+Model was trained by Auri.
+
+Training notes
+
+This model was trained for 2 epochs on 10k rows (~18.7M tokens), taken equally from Erebus-87k and r_shortstories_24k datasets. I've also normalized punctuation to ASCII on the train split, so mismatched quote marks should not be an issue anymore. Also normalized whitespaces, so double spaces after period should be gone as well.
+
+It was trained on 5x3090Ti workstation for 7.5 hours with rsLoRA. I switched back to Axolotl for this run, as LF just plain refused to run at all on this workstation. Also, it's a bf16 LoRA this time. Overall training went much smoother than last time. I've attempted to train Qwen Sugarquill several times before, but loss jumped like crazy. Effective batch size of 40, rsLoRA and paged_ademamix_8bit optimizer seemingly completely solved this issue.
+
+Thanks to Kearm for providing compute for this training run!
+
+Format
+
+Model responds to ChatML instruct formatting, exactly like it's base model.
+
+<|im_start|>system
+{system message}<|im_end|>
+<|im_start|>user
+{user message}<|im_end|>
+<|im_start|>assistant
+{response}<|im_end|>
+
+Recommended Samplers
+
+I found this configuration to be quite stable:
+
+Temperature - 0.8
+Min-P - 0.05
+Top-A - 0.3
+Repetition Penalty - 1.03
+
+Feel free to toy around with samplers after you get a feel for it. It seems to like Top-A and Smooth Sampling quite a bit.
+
+Training config
+See Axolotl config
+
+axolotl version: 0.4.1
+
+# Model
+base_model: arcee-ai/SuperNova-Medius
+strict: false
+
+# Liger Kernels (optimization)
+plugins:
+  - axolotl.integrations.liger.LigerPlugin
+liger_rope: true
+liger_rms_norm: true
+liger_swiglu: true
+liger_fused_linear_cross_entropy: true
+
+# Output and HuggingFace
+output_dir: /home/kearm/axolotl/TQ-2.5-14B-Sugarquill
+hub_model_id: allura-org/TQ-2.5-14B-Sugarquill-LoRA
+hf_use_auth_token: true
+hub_strategy: "all_checkpoints"
+
+# WandB
+wandb_project: huggingface
+wandb_entity:
+wandb_name: TQ-2.5-14B-Sugarquill-1
+
+# Data
+#chat_template: chatml
+#train_on_inputs: false
+group_by_length: false
+datasets:
+  - path: allura-org/sugarquill-10k
+    type: completion
+
+## Evaluation
+val_set_size: 0.01
+evals_per_epoch: 4
+eval_table_size:
+eval_max_new_tokens: 128
+
+# Technical aspects
+sequence_len: 8192
+save_safetensors: true
+saves_per_epoch: 2
+logging_steps: 1
+special_tokens:
+
+# Quantization
+bf16: auto
+fp16:
+tf32: false
+## For LoRA
+load_in_8bit: false
+load_in_4bit: false
+
+# LoRA
+peft_use_rslora: true
+peft_use_dora: false # better but slower
+adapter: lora # lora or qlora
+lora_model_dir:
+lora_r: 64 # 64 is optimal for most trains on instruct
+lora_alpha: 32
+lora_dropout: 0.1
+lora_target_linear: true
+lora_fan_in_fan_out:
+lora_target_modules:
+#  - embed_tokens
+#  - lm_head
+
+#loraplus_lr_ratio: 8 # works to converge faster but is kinda cancer bc makes model unstable
+#loraplus_lr_embedding:
+
+# Training hyperparameters
+# max_steps:
+num_epochs: 2
+
+# Anti Overfit and Stability
+weight_decay: 0.01
+max_grad_norm: 1.0
+
+## Learning Rate
+warmup_ratio: 0.05
+learning_rate: 0.00003
+lr_scheduler: cosine
+#lr_scheduler_kwargs:
+#    min_lr: 0.0000024
+optimizer: paged_ademamix_8bit # usually adamw_torch or paged_adamw_8bit
+
+## Batch Size
+gradient_accumulation_steps: 8      # More effective batch size - stabler train, usually. MBS also speeds it up.
+micro_batch_size: 1                 # Batch size per gpu = micro_batch_size * gradient_accumulation_steps
+eval_batch_size: 1
+
+# Optimizations
+pad_to_sequence_len: true
+sample_packing: true
+eval_sample_packing: false
+flash_attention: true
+xformers_attention:
+gradient_checkpointing: "unsloth"
+gradient_checkpointing_kwargs:
+   use_reentrant: true
+local_rank:
+deepspeed: /home/kearm/axolotl/deepspeed_configs/zero3_bf16.json # Only use with multi gpu # _bf16_cpuoffload_all
+# fsdp:
+#   - full_shard
+#   - auto_wrap
+# fsdp_config:
+#   fsdp_limit_all_gathers: true
+#   fsdp_sync_module_states: true
+#   fsdp_offload_params: true
+#   fsdp_use_orig_params: false
+#   fsdp_cpu_ram_efficient_loading: true
+#   fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
+#   fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer
+#   fsdp_state_dict_type: FULL_STATE_DICT
+#   fsdp_sharding_strategy: FULL_SHARD
+# Misc
+early_stopping_patience:
+debug:
+
+---
 ## Use with llama.cpp
 Install llama.cpp through brew (works on Mac and Linux)