♟️ Model Card — Mattimax/EliaChess-70M

☕ Support my research

📌 Overview

Mattimax/EliaChess-70M is a ~70 million parameter Small Language Model (SLM) designed to generate chess moves for agentic gameplay within chess arenas, enabling autonomous competition against other LLMs. Trained on the mlabonne/chessllm dataset, it focuses on producing valid and context-aware moves in standard chess notation, optimizing for interaction-driven play rather than deep engine-level analysis.

• Author: Mattimax
• Model Name: EliaChess-70M
• Parameters: ~70M
• Architecture: Transformer (decoder-only)
• Frameworks: PyTorch, Hugging Face Transformers
• Primary Language: English (with some generalization capability)
• License: (to be specified)

---

🎯 Intended Use

This model is built for:

• Chess move generation (basic/intermediate level)
• Position analysis and explanation
• Educational support for chess learners
• Conversational chess assistants
• Lightweight AI applications (local inference, edge devices)

---

📚 Training Data

The model was trained using:

• Dataset: mlabonne/chessllm
• Structured chess data including:
  • PGN game records
  • Opening sequences
  • Move annotations
• Synthetic data augmentation to improve robustness
• Additional general NLP data to enhance fluency

Preprocessing

• Tokenization via Transformers tokenizer
• Standardization of chess notation (SAN, PGN, FEN)
• Cleaning and filtering of noisy samples

---

⚙️ Capabilities

✔️ Strengths

• Understands standard chess notation (e.g., e4, Nf3, O-O)
• Explains basic strategies and concepts
• Generates plausible moves in simple positions
• Maintains coherent chess-related conversations

⚠️ Limitations

• Not a substitute for advanced engines like Stockfish
• Limited tactical depth and calculation ability
• May produce illegal or suboptimal moves
• Struggles with complex or deep positions

---

🧠 Architecture & Training

• Type: Decoder-only Transformer
• Size: ~70M parameters
• Framework: PyTorch + Hugging Face Transformers
• Training Approach: Fine-tuning on domain-specific data

Hyperparameters (approximate)

• Learning Rate: 5e-5 – 1e-4
• Batch Size: Variable (hardware dependent)
• Epochs: 3–10
• Context Length: 512–2048 tokens

---

🧪 Evaluation

Evaluated on:

• Move prediction tasks
• Chess-related Q&A
• Position understanding

Observations

• Good performance on beginner/intermediate tasks
• Strong linguistic coherence
• Limited strategic depth

---

🚀 Use Cases

• Chess learning assistants
• Embedded AI in chess apps
• Chat-based chess tools
• Research on domain-specific SLMs
• Local/offline AI systems

---

⚠️ Ethical Considerations

• Outputs may be incorrect or misleading
• Should not be used in competitive environments without validation
• Users should verify critical outputs with reliable engines

---

🔧 Integration

Compatible with:

• Hugging Face Transformers
• Local inference pipelines
• Ollama (after conversion)

📈 Roadmap

• Improve tactical reasoning
• Expand and refine training dataset the
• Hybrid integration with chess engines

---

👤 Credits

Developed by Mattimax, focused on efficient and specialized AI systems.