llama-3.2-1b-math-solver

LLaMA-3.2-1B-Math (QLoRA)

Model Summary

LLaMA-3.2-1B-Math is a compact math-reasoning model fine-tuned from LLaMA 3.2-1B using 4-bit QLoRA. It is designed to study domain-specialized adaptation under strict compute limits.

Model Details

• Base model: LLaMA 3.2-1B
• Architecture: Decoder-only Transformer
• Fine-tuning: QLoRA (4-bit NF4)
• Training environment: Kaggle (single-GPU)

Datasets

Datasets were processed into a plain-text format (no chat template) to match the base model.

Training Procedure

• LoRA adapters inserted into attention and MLP projections
• Mixed math datasets concatenated and split into train/eval
• Optimized for stability on small batch sizes

Intended Use

• Math reasoning benchmarks
• Research on small-model specialization
• Educational demonstrations of QLoRA fine-tuning

Out-of-Scope Uses

• General chat or instruction following
• Safety-critical or production systems

Known Limitations

• Struggles with long proofs and symbolic manipulation
• Sensitive to prompt phrasing
• No RLHF or instruction tuning applied

Ethical Considerations

This model inherits biases and limitations from its base model and training data.

Acknowledgements

• LLaMA (Meta)
• GSM8K (OpenAI)
• MATH (fdyrd/MATH)