llm-tool-call-lora-Qwen0.5B

🛠️ Qwen 2.5-0.5B Tool Call Fine-Tuning (Phase 1 — Synthetic SFT)

Fine-tuning Qwen 2.5-0.5B on a small synthetic dataset to make structured JSON tool calls using QLoRA.

This is Phase 1 of a 3-phase experiment. See the full experiment repo for SFT + GRPO on real data.

---

📌 What This Does

Trains a 0.5B model to respond to natural language with structured tool calls:

{"tool": "get_stock_price", "arguments": {"ticker": "MSFT"}}

---

🤖 Setup

---

📊 Results

Evaluated on 10 queries (5 seen during training, 5 unseen):

Key observations:

• Model learned the concept of JSON tool calls but hallucinated tool names ("weather" instead of "get_weather")
• Wrong argument keys ("ticker_symbol" instead of "ticker")
• Calculator answered directly in plain text instead of calling the tool
• 14 examples is too small — model memorises poorly and doesn't generalise

---

📁 Files

---

💡 Lessons Learned

• 14 examples is not enough — model fails to generalise even to similar queries
• Exact argument matching is strict — "ticker_symbol" vs "ticker" counts as failure
• Bigger dataset + RL needed → See Phase 2 & 3 for improvement

---

🔧 Stack

Qwen2.5-0.5B • QLoRA • SFT • trl • peft • Google Colab T4