gpt2-small_instruction

Fine-Tuned GPT-2 Instruction Follower (Alpaca)

This release contains the fine-tuned model weights (gpt2-small_instruction.pth) for the Instruction Following task. These weights allow users to run inference or evaluation immediately without needing to retrain the model from scratch.

---

Performance Metrics

The model was fine-tuned on a lightweight subset (5,000 examples) of the official Stanford Alpaca dataset for pipeline verification. It was evaluated using an automated LLM-as-a-Judge (gemma4:e2b) on a holdout test set.

Metric	Value
LLM Judge Score	39.59 / 100
Training Samples	~4,250 (85% of 5k subset)
Base Model	GPT-2 Small (124M)

---

Usage Instructions

1. Download gpt2-small_instruction.pth from the Assets section below.
2. Place the file in the instruction-finetuning/model_weights/ directory.
3. Run the main.ipynb notebook. Set MODE = "inference" and execute the Interactive Generation cells. The pipeline will detect the weights:

   # Expected output:
   --- PREPARING FOR INFERENCE ---
   Loading finetuned model from: ./model_weights/gpt2-small_instruction.pth
   Moving model to mps...
   

Critical Requirement:
These weights are strictly coupled to the GPT-2 Small (124M) architecture.

• Do not try to load these into gpt2-medium or other sizes.
• Do not try to load these into a standard Hugging Face model class directly; they require the custom GPTModel architecture defined in our root src.model.

---

Generated by the instruction-finetuning/main.ipynb pipeline on Apple Silicon (M3 Pro)