Updated README

README.md

@@ -0,0 +1,88 @@
+---
+license: apache-2.0
+datasets:
+- ise-uiuc/Magicoder-OSS-Instruct-75K
+- ise-uiuc/Magicoder-Evol-Instruct-110K
+language:
+- en
+- zh
+- ja
+- de
+- es
+base_model:
+- Qwen/Qwen3.5-9B
+tags:
+- gcm
+- qwen3_5
+- reasoning
+- qwen
+---
+
+<div align="center">
+  <img src="https://cdn-uploads.huggingface.co/production/uploads/69c842686cf758859915159c/EfCpXkcGtQhv02oKAy2La.png" width="700">
+</div>
+
+# GCM Mark II
+ 
+GCM Mark II is a QLoRA fine-tune of **Qwen3.5-9B**, trained to improve coding reliability — specifically constraint-following, edge-case handling, and reducing invented/hallucinated API usage.
+
+<div align="center">
+  <img src="https://cdn-uploads.huggingface.co/production/uploads/69c842686cf758859915159c/L4Aa_zxIgHRi_qkIsGrjP.png" width="700">
+</div>
+
+ 
+## Model Details
+ 
+- **Base model:** Qwen3.5-9B
+- **Fine-tuning method:** QLoRA & CPT
+- **Tokens trained:** ~2.5 Million
+- **Training data:** [`ise-uiuc/Magicoder-Evol-Instruct-110K`](https://huggingface.co/datasets/ise-uiuc/Magicoder-Evol-Instruct-110K) (partial epoch)
+- **License:** Apache 2.0
+ 
+## Intended Use
+ 
+General-purpose code generation and coding assistance across multiple languages (Python, JavaScript, Go, C, C++, Java, Rust tested directly). Not evaluated for production/safety-critical code without independent review — see Known Limitations below.
+ 
+## Benchmark Results
+ 
+Evaluated on a custom 20-question coding eval (mixed languages, weighted toward hard/edge-case problems), hand-graded, single greedy pass, thinking mode enabled. Full methodology, per-question breakdown, and comparison against OpenGCM-v2 and base Qwen3.5-9B are in [`BENCHMARK.md`](./BENCHMARK.md) in this repo.
+ 
+| Model | Score |
+|---|---|
+| OpenGCM-v2 | 17/20 (85%) |
+| **GCM Mark II** | **16/20 (80%)** |
+| Qwen3.5-9B (base) | 6/20 (30%)* |
+ 
+*Base model score includes multiple non-completions rather than purely incorrect answers — see `BENCHMARK.md` for details.
+ 
+This comparison is reported as-is, including the result where a competing model scored higher, because the point of publishing it is to be checkable, not to win. n=20 is a small sample — treat these numbers as directional, not definitive, and feel free to reproduce or challenge them.
+ 
+**Qualitative note:** GCM Mark II generated responses quickly and reliably across all 20 questions, with no incomplete generations or stuck reasoning loops — worth weighing alongside the raw accuracy numbers if generation reliability matters for your use case.
+ 
+## Known Limitations
+ 
+- **Directed graph algorithms:** GCM Mark II has a specific, reproducible weakness in cycle detection on directed graphs — it can conflate "visited" with "currently on the active recursion path," causing false-positive cycle detection on some acyclic graphs (e.g. diamond-shaped DAGs). If you're using this model for graph algorithms, verify output independently.
+- Small eval sample (n=20) — broader capability outside the tested question set is not guaranteed.
+- Not evaluated against standardized public benchmarks (HumanEval, LiveCodeBench, MBPP) yet — evaluation attempted but blocked by local tooling issues during development; may be added in a future update.
+
+## How to Use
+ 
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+ 
+model = AutoModelForCausalLM.from_pretrained("OPENGCM/GCM-MARK-II", device_map="auto")
+tokenizer = AutoTokenizer.from_pretrained("OPENGCM/GCM-MARK-II")
+ 
+messages = [{"role": "user", "content": "Write a function to check if a binary tree is balanced."}]
+inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
+outputs = model.generate(inputs, max_new_tokens=1024)
+print(tokenizer.decode(outputs[0], skip_special_tokens=True))
+```
+ 
+## Citation / Attribution
+ 
+Base model: Qwen3.5-9B (Qwen team).
+Training data: Magicoder-Evol-Instruct-110K (ise-uiuc).
+
+# Ollama / GGUF Support
+OpenGCM is actively working on .gguf files for quantized versions of GCM Mark II. Stay tuned!