chess-tinyllama

♟️ Konvah's Chess TinyLlama

This model is a fine-tuned version of TinyLlama/TinyLlama-1.1B-Chat-v1.0 using LoRA for the Aura Chess AI Tournament. It predicts high-quality chess moves in JSON format, given a move history, color, and a list of legal moves.

---

🧠 Model Objective

The model learns to:

• Choose the best legal move (move)
• Give a short explanation (reasoning) in ≤10 words
• Format responses as valid JSON
• Respond in [INST] ... [/INST] format

---

💡 Input Format

The model uses structured prompts:

[INST]
You are a chess player.
{"moveHistory": ["e4", "e5", "Nf3"], "possibleMoves": ["Nc3", "Bc4", "d4"], "color": "w"}
[/INST]

🎯 Output Format
Always a single-line JSON:

json
Copy
Edit
{"move": "Bc4", "reasoning": "Develops bishop and targets f7"}
The move must be from possibleMoves

The reasoning is free-form but short

🛠️ Training Details
Base: TinyLlama-1.1B-Chat

LoRA (8-bit): q_proj, k_proj, v_proj, o_proj

Epochs: 3

Dataset: ~70 samples from master-level PGNs

Format: instruction-style using transformers.Trainer

📈 Performance
| Metric      | Value |
| ----------- | ----- |
| Final loss  | 1.08  |
| Epochs      | 3     |
| Batch size  | 1     |
| Total steps | 51    |

🚀 Usage
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("Konvah/chess-tinyllama")
tokenizer = AutoTokenizer.from_pretrained("Konvah/chess-tinyllama")

prompt = """[INST]
You are a chess player.
{"moveHistory": ["e4", "e5", "Nf3"], "possibleMoves": ["Nc3", "Bc4", "d4"], "color": "w"}
[/INST]"""

inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=50)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

📎 License
Open for research and tournament evaluation. Not intended for production without additional safety testing.

✍️ Author
Ismail Abubakar (@boringcrypto_)

Contact: [email protected]

🏆 Aura Tournament
This model was created for the Aura Chess LLM Tournament to demonstrate reasoning and strategy prediction using open-source LLMs.

---