Create README.md

README.md

@@ -0,0 +1,81 @@
+---
+language: en
+license: gemma
+tags:
+  - gemma-3
+  - chess
+  - instruction-tuning
+  - sft
+  - lora
+  - merged
+  - unsloth
+  - trl
+  - small-model
+base_model: unsloth/gemma-3-270m-it
+library_name: transformers
+pipeline_tag: text-generation
+datasets:
+  - Thytu/ChessInstruct
+model-index:
+  - name: Finetuned-gemma3-270m-chess-merged
+    results: []
+---
+
+# Gemma-3 270M — Chess Coach (Merged FP16)
+
+**Author:** [@codertrish](https://huggingface.co/codertrish)  
+**Base model:** [`unsloth/gemma-3-270m-it`](https://huggingface.co/unsloth/gemma-3-270m-it)  
+**Type:** *Merged* FP16 checkpoint (LoRA deltas baked into base — no adapters needed)  
+**Task:** Conversational chess tutoring (rules, beginner principles, simple reasoning)
+
+---
+
+## TL;DR
+
+- This is a **plug-and-play** Gemma-3 (270M) checkpoint specialized for **chess coaching**.  
+- It was fine-tuned via **LoRA** on a subset of **`Thytu/ChessInstruct`**, then **merged to FP16**.  
+- Load directly with `transformers` and chat using the **Gemma-3 chat template**.
+
+---
+
+## ✨ Intended Use
+
+- **Direct use:** Explain chess rules, beginner opening principles, basic tactics, and high-level strategy in **plain text**.  
+- **Downstream use:** As a small assistant embedded in notebooks, tutorials, or beginner-level chess learning tools.
+
+**Out-of-scope:** Engine-level move search, advanced calculation, or authoritative evaluations of complex positions. For serious analysis, use a dedicated chess engine (e.g., Stockfish) and verify claims.
+
+---
+
+## 🔧 How to Use
+
+> The model expects **Gemma-3 chat formatting**. Use `apply_chat_template` before generation.
+
+### Minimal example (Transformers)
+```python
+from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline
+
+REPO = "codertrish/Finetuned-gemma3-270m-chess-merged"
+
+tok   = AutoTokenizer.from_pretrained(REPO)
+model = AutoModelForCausalLM.from_pretrained(REPO, torch_dtype="bfloat16", device_map="auto")
+
+pipe = pipeline("text-generation", model=model, tokenizer=tok, return_full_text=False)
+
+def chat(messages, **gen_kwargs):
+    prompt = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+    eot_id = tok.convert_tokens_to_ids("<end_of_turn>")
+    out = pipe(
+        prompt,
+        eos_token_id=eot_id,
+        max_new_tokens=200,
+        do_sample=False,         # deterministic; set True for sampling
+        **gen_kwargs,
+    )[0]["generated_text"]
+    return out.strip()
+
+messages = [
+    {"role":"system","content":"You are a helpful chess coach. Answer in plain text, 3 concise bullets."},
+    {"role":"user","content":"What are the main opening principles?"},
+]
+print(chat(messages))