jondurbin/airoboros-l2-7b-gpt4-2.0-tuning.md

## airoboros-l2-7b-gpt4-2.0-tuning.md

      
    Raw
  

              airoboros-l2-7b-gpt4-2.0-tuning.md
            
          
    Overview

This was a full fine-tune of llama-2-7b-hf using dataset https://huggingface.co/datasets/jondurbin/airoboros-gpt4-2.0
Data prep

Convert the JSONL (newline delimeted JSON strings) into conversational format that FastChat expects:
import re
import json
import uuid
inputs = [json.loads(line) for line in open("instructions.jsonl").readlines()]

def split_response(instruction, response):
    if '</s>' not in response:
        return [
            {
                "from": "human",
                "value": instruction,
            },
            {
                "from": "gpt",
                "value": response,
            },
        ]

    parts = response.split('</s>')
    user = [instruction]
    assistant = []
    for idx in range(len(parts)):
        part = parts[idx]
        if idx == 0:
            assistant.append(part)
            continue
        match = re.match(r'^\s*USER:(.*?)ASSISTANT:(.*)\s*$', part, re.DOTALL)
        if not match:
            return None
        user.append(match.group(1).strip())
        assistant.append(match.group(2).strip())
    conv = []
    for idx in range(len(user)):
        conv.append({
            "from": "human",
            "value": user[idx],
        })
        conv.append({
            "from": "gpt",
            "value": assistant[idx]
        })
    return conv

conversations = []
for row in inputs:
    conversation = split_response(row['instruction'], row['response'])
    if not conversation:
        print("Bad format, skipping...")
        continue
    conversations.append({
        "id": str(uuid.uuid4()),
        "conversations": conversation,
    })
with open("as_conversations.json", "w") as outfile:
    outfile.write(json.dumps(conversations, indent=2))
with FastChat (and flash attention 2.0.1)
Small patch to FastChat

Need to use the airoboros format instead of vicuna, plus use trainer.save_model:
diff --git a/fastchat/train/train.py b/fastchat/train/train.py
index 0fa855e..948bac3 100644
--- a/fastchat/train/train.py
+++ b/fastchat/train/train.py
@@ -82,7 +82,7 @@ def preprocess(
     sources,
     tokenizer: transformers.PreTrainedTokenizer,
 ) -> Dict:
-    conv = get_conversation_template("vicuna")
+    conv = get_conversation_template("airoboros_v1")
     roles = {"human": conv.roles[0], "gpt": conv.roles[1]}

     # Apply prompt templates
@@ -272,7 +272,8 @@ def train():
         trainer.train()
     model.config.use_cache = True
     trainer.save_state()
-    safe_save_model_for_hf_trainer(trainer=trainer, output_dir=training_args.output_dir)
+    trainer.save_model(output_dir=training_args.output_dir)
+    #safe_save_model_for_hf_trainer(trainer=trainer, output_dir=training_args.output_dir)


 if __name__ == "__main__":

Hardware

This was done with 4x 80GB A100 GPUs.
Deepspeed config

{
  "gradient_accumulation_steps": "auto",
  "gradient_clipping": "auto",
  "train_batch_size": "auto",
  "train_micro_batch_size_per_gpu": "auto",
  "bf16": {
    "enabled": true
  },
  "zero_optimization": {
    "stage": 2,
    "contiguous_gradients": true,
    "overlap_comm": true,
    "reduce_scatter": true,
    "reduce_bucket_size": 5e8,
    "allgather_bucket_size": 5e8
  }
}
Fine-tune

export WANDB_API_KEY=[redacted]
export BASE_MODEL=llama-2-7b-hf
export WANDB_PROJECT=airoboros-l2-7b-gpt4-m2.0
export PYTHONPATH=FastChat

deepspeed --num_nodes=1 --num_gpus=4 FastChat/fastchat/train/train_mem.py \
  --model_name_or_path $BASE_MODEL \
  --data_path as_conversations.json \
  --output_dir $WANDB_PROJECT \
  --num_train_epochs 3 \
  --evaluation_strategy "no" \
  --save_steps 100 \
  --save_total_limit 1 \
  --learning_rate 2e-5 \
  --weight_decay 0. \
  --warmup_ratio 0.04 \
  --lr_scheduler_type "cosine" \
  --logging_steps 1 \
  --bf16 \
  --per_device_train_batch_size 12 \
  --deepspeed deepspeed.json \
  --model_max_length 4096 \
  --gradient_checkpointing True \
  --lazy_preprocess True