Nanosft

A minimal supervised fine-tuning (SFT) setup for QnA-style dialogue models, building upon Karpathy’s nanoGPT.
This project demonstrates how to implement prompt masking and a chat-style template for fine-tuning small GPT models on conversational datasets such as Databricks Dolly-15k.

---

Features

• QnA prompt–response formatting (User: / Assistant: template)
• Label masking to train only on assistant responses
• Simple PyTorch + Datasets data pipeline
• Configurable logging with Weights & Biases
• AMP training support for faster convergence

---

Usage

python nanosft.py \
  --dataset databricks/databricks-dolly-15k \
  --hf_model_type gpt2 \
  --batch_size 8 \
  --grad_accum 4 \
  --epochs 1 \
  --lr 3e-5 \
  --use_amp

---

Future Work

• Optimization for larger-scale runs (gradient checkpointing, mixed precision)
• Multi-GPU and distributed training support
• LoRA-based fine-tuning for efficient parameter updates

Note: currently only a educational version, not prod ready