PursuitOfDataScience/llama3.2-1b-dpo

This repository contains a two-stage fine-tuned version of meta-llama/Llama-3.2-1B:

1. Supervised fine-tuning (SFT) on a local copy of HuggingFaceH4/ultrachat_200k using an instruction-style, multi-turn chat objective.
2. Direct Preference Optimization (DPO) on a local copy of mlabonne/orpo-dpo-mix-40k, using preference pairs derived from instruction-following data.

Model details

• Base model: meta-llama/Llama-3.2-1B
• Stage 1 objective: Supervised fine-tuning for helpful, concise chat responses on Ultrachat-style conversations.
• Stage 2 objective: DPO on preference pairs (chosen vs. rejected responses) from mlabonne/orpo-dpo-mix-40k.
• Context length: Up to 4096 tokens (subject to the base model config).
• Training data:
  • SFT: multi-turn dialogues from HuggingFaceH4/ultrachat_200k.
  • DPO: preference pairs from mlabonne/orpo-dpo-mix-40k.

Inference usage

The model is trained in a chat-style setup. At inference time, prompts are built as a list of messages and passed through the model's native chat_template via tokenizer.apply_chat_template:

from transformers import AutoTokenizer, AutoModelForCausalLM

repo_id = "PursuitOfDataScience/llama3.2-1b-dpo"

tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForCausalLM.from_pretrained(
    repo_id,
    device_map="auto",
)

messages = [
    {
        "role": "system",
        "content": (
            "You are a helpful, concise assistant. "
            "Write clear, well-structured answers that follow the user's constraints."
        ),
    },
    {
        "role": "user",
        "content": "Explain how someone can build a consistent daily learning habit.",
    },
]

prompt_text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)

inputs = tokenizer(prompt_text, return_tensors="pt").to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=512,
    eos_token_id=tokenizer.eos_token_id,
    pad_token_id=tokenizer.eos_token_id,
    temperature=0.7,
    top_p=0.9,
    do_sample=True,
)

# Decode only the generated continuation (excluding the prompt tokens)
generated_tokens = outputs[0][inputs["input_ids"].shape[1]:]
response = tokenizer.decode(generated_tokens, skip_special_tokens=True)
print(response)

Multi-turn example

messages = [
    {
        "role": "system",
        "content": (
            "You are a helpful, concise assistant. "
            "Write clear, well-structured answers that follow the user's constraints."
        ),
    },
    {
        "role": "user",
        "content": "Describe the main trade-offs between using small and large language models.",
    },
    {
        "role": "assistant",
        "content": "Small models are cheaper and faster, while large models are usually more capable...",
    },
    {
        "role": "user",
        "content": "Give me a bullet-point summary from the perspective of a startup.",
    },
]

prompt_text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)

inputs = tokenizer(prompt_text, return_tensors="pt").to(model.device)
outputs = model.generate(
    **inputs,
    max_new_tokens=256,
    eos_token_id=tokenizer.eos_token_id,
    pad_token_id=tokenizer.eos_token_id,
    temperature=0.7,
    top_p=0.9,
    do_sample=True,
)
response = tokenizer.decode(
    outputs[0][inputs["input_ids"].shape[1]:],
    skip_special_tokens=True,
)
print(response)

Training pipeline (summary)

1. Instruction SFT (Ultrachat):

   • Conversations are converted into lists of messages.
   • For each assistant turn, a single training example is built using tokenizer.apply_chat_template.
   • Loss is applied only on assistant tokens; system and user tokens are masked.

2. DPO finetuning (ORPO DPO Mix):

   • Preference pairs (chosen/rejected responses) are pre-processed using the same chat template logic to ensure consistency with inference.
   • The DPO objective is applied using TRL's DPOTrainer, with prompts and chosen/rejected continuations derived from the pre-tokenized data.
   • See the dpo.py script used in this project for full configuration (batch sizing, max length, etc.).

Limitations

• This is a relatively small (1B parameter) model and may hallucinate or struggle on complex, multi-step reasoning tasks.
• Outputs may be inaccurate, unsafe, or biased. Always verify critical information before using it in production.