Model Card for pricer-2025-12-01_15.09.36

This adapter fine-tunes meta-llama/Meta-Llama-3.1-8B with LoRA/PEFT to predict product prices from textual descriptions. Training followed the setup described in this repo’s README.md using TRL and the costadev00/pricer-data dataset.

How to use

from transformers import pipeline

prompt = (
    "You are a pricing assistant. "
    "Product description: Apple iPhone 15 Pro, 256GB, new in box.\n"
    "Price is $"
)

pipe = pipeline(
    "text-generation",
    model="costadev00/pricer-2025-12-01_15.09.36",
    device_map="auto",
)
result = pipe(
    [{"role": "user", "content": prompt}],
    max_new_tokens=16,
    return_full_text=False,
)
print(result[0]["generated_text"])

Note: load the base model plus LoRA adapter via peft if you downloaded the adapter-only weights.

Intended use & limitations

• Use cases: price suggestion or ranking experiments for e-commerce catalogs similar to the training data.
• Limitations: the model’s accuracy degrades for domains or currencies not covered in the dataset; treat outputs as hints, not ground truth.
• Ethical considerations: do not deploy without safeguards against hallucinations or unfair pricing decisions.

Training procedure

• Data: costadev00/pricer-data (Amazon-style product descriptions + prices) split into train/test as hosted on the HF Hub.
• Hardware: single 80 GB NVIDIA A100.
• Method: QLoRA (4-bit) + SFT via trl.SFTTrainer.
• Hyperparameters:
  • epochs: 2
  • batch size: 16 (per device)
  • grad accumulation: 1
  • learning rate: 1e-4 (cosine schedule, warmup 3%)
  • LoRA: r=32, alpha=64, dropout=0.1, target modules q_proj,k_proj,v_proj,o_proj
  • optimizer: paged_adamw_32bit
  • max sequence length: 182 tokens
  • checkpointing: every 2 000 steps, pushed privately to the Hub

Framework versions

• TRL: 0.14.0
• Transformers: 4.48.3
• PyTorch: 2.1.x (CUDA 12.4 build)
• Datasets: 3.2.0
• Tokenizers: 0.21.4
• BitsAndBytes: 0.46.0
• PEFT: 0.14.0

Evaluation

• Metric: MAE on held-out product descriptions (lower is better).
• Result: Add your value here once measured.
• Include qualitative inspection of generated prices before deploying.

Citation

If you use this model or training recipe, please cite TRL:

@misc{costadev00_llama3_price_predictor_lora_2025,
  author       = {MONTEIRO, Matheus Costa},
  title        = {Llama-3 Price Predictor Fine-Tuning (LoRA)},
  year         = {2025},
  publisher    = {GitHub},
  journal      = {GitHub repository},
  howpublished = {\url{https://github.com/costadev00/fine-tuning-llama3-8b}},
  note         = {Accessed: 2025-12-29}
}

License

Replace license above with the actual license governing your fine-tuned weights (e.g., apache-2.0, llama3-community, or custom terms). Ensure compliance with Meta’s Llama 3.1 usage policy.