flammen24-mistral-7B

A Mistral 7B LLM built from merging pretrained models and finetuning on jondurbin/truthy-dpo-v0.1. Flammen specializes in exceptional character roleplay, creative writing, and general intelligence

Method

Finetuned using an A100 on Google Colab.

Fine-tune a Mistral-7b model with Direct Preference Optimization - Maxime Labonne

Configuration

System prompt, dataset formatting:

def chatml_format(example):

 # Format system
 system = ""
 if example.get('system') and len(example['system']) > 0:
 systemMessage = example['system']
 system = "<|im_start|>system\n" + systemMessage + "<|im_end|>\n"

 # Format instruction
 prompt = "<|im_start|>user\n" + example['prompt'] + "<|im_end|>\n<|im_start|>assistant\n"

 # Format chosen answer
 chosen = example['chosen'] + "<|im_end|>\n"

 # Format rejected answer
 rejected = example['rejected'] + "<|im_end|>\n"

 return {
 "prompt": system + prompt,
 "chosen": chosen,
 "rejected": rejected,
 }

dataset = load_dataset("jondurbin/truthy-dpo-v0.1")['train']

# Save columns
original_columns = dataset.column_names

# Tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_name)
tokenizer.pad_token = tokenizer.eos_token
tokenizer.padding_side = "left"

# Format dataset
dataset = dataset.map(
 chatml_format,
 remove_columns=original_columns
)

LoRA, model, and training settings:

# LoRA configuration
peft_config = LoraConfig(
 r=16,
 lora_alpha=16,
 lora_dropout=0.05,
 bias="none",
 task_type="CAUSAL_LM",
 target_modules=['k_proj', 'gate_proj', 'v_proj', 'up_proj', 'q_proj', 'o_proj', 'down_proj']
)

# Model to fine-tune
model = AutoModelForCausalLM.from_pretrained(
 model_name,
 torch_dtype=torch.bfloat16,
 load_in_4bit=True
)
model.config.use_cache = False

# Reference model
ref_model = AutoModelForCausalLM.from_pretrained(
 model_name,
 torch_dtype=torch.bfloat16,
 load_in_4bit=True
)

# Training arguments
training_args = TrainingArguments(
 per_device_train_batch_size=4,
 gradient_accumulation_steps=4,
 gradient_checkpointing=True,
 learning_rate=5e-5,
 lr_scheduler_type="cosine",
 max_steps=200,
 save_strategy="no",
 logging_steps=1,
 output_dir=new_model,
 optim="paged_adamw_32bit",
 warmup_steps=100,
 bf16=True,
 report_to="wandb",
)

# Create DPO trainer
dpo_trainer = DPOTrainer(
 model,
 ref_model,
 args=training_args,
 train_dataset=dataset,
 tokenizer=tokenizer,
 peft_config=peft_config,
 beta=0.1,
 max_prompt_length=2048,
 max_length=8192,
 force_use_ref_model=True
)