Upload README.md with huggingface_hub

README.md

@@ -0,0 +1,69 @@
+---
+license: apache-2.0
+base_model: 01-ai/Yi-Coder-9B-Chat
+language:
+  - en
+tags:
+  - code
+  - coding-assistant
+  - gguf
+  - ollama
+  - qlora
+  - yi-coder
+pipeline_tag: text-generation
+library_name: gguf
+---
+
+# AXIOM-coder-9B
+
+**AXIOM** is a private, open coding assistant built by **Leslie Moody** — a [Yi-Coder 9B](https://huggingface.co/01-ai/Yi-Coder-9B-Chat) model fine-tuned (QLoRA) on Leslie's own code plus a Python coding-instruction dataset, then merged and quantized to **GGUF (Q4_K_M)** so it runs fully **locally and privately** through Ollama or llama.cpp.
+
+It identifies itself as *AXIOM, built by Leslie Moody*, focuses on **coding** (writing, explaining, debugging, refactoring), and is tuned to be honest and grounded.
+
+## Quick start (Ollama)
+
+1. Download `axiom-v1.Q4_K_M.gguf` and `Modelfile` from this repo into the same folder.
+2. Create and run the model:
+
+```bash
+ollama create axiom:v1 -f Modelfile
+ollama run axiom:v1 "Write a Python function that reverses a linked list, with a docstring."
+```
+
+The included `Modelfile` carries AXIOM's system prompt (identity + coding focus + honesty) and sensible parameters (`temperature 0.25`, `num_ctx 8192`).
+
+### llama.cpp
+
+```bash
+llama-cli -m axiom-v1.Q4_K_M.gguf -p "Refactor this function for readability: ..."
+```
+
+## Files
+
+| File | Size | What |
+|------|------|------|
+| `axiom-v1.Q4_K_M.gguf` | ~5.3 GB | the quantized model (Q4_K_M) |
+| `Modelfile` | tiny | Ollama Modelfile with AXIOM's system prompt + params |
+
+## Training
+
+- **Base:** `01-ai/Yi-Coder-9B-Chat` (Apache 2.0)
+- **Method:** QLoRA (4-bit), rank 16, 1 epoch
+- **Data (6,964 examples):** ~1,964 docstring↔code pairs extracted from Leslie's own Python projects + 5,000 rows of [`iamtarun/python_code_instructions_18k_alpaca`](https://huggingface.co/datasets/iamtarun/python_code_instructions_18k_alpaca)
+- **Export:** merged to fp16 → converted to GGUF → quantized to Q4_K_M (llama.cpp)
+- Trained on a single free-tier GPU.
+
+## Intended use
+
+A personal / local coding assistant you fully own and run offline. Great for code help, explanations, and quick refactors.
+
+## Limitations (read this)
+
+- It's a **9B fine-tune, not a frontier model.** It won't match large hosted assistants on hard, long-context, or agentic coding tasks.
+- **English + code focused.** Non-English ability is limited (the base model isn't broadly multilingual).
+- Like all LLMs it can **be confident and wrong** — verify generated code, and never rely on it for safety- or security-critical decisions without review.
+- A small model fine-tuned for 1 epoch can occasionally over-fit phrasing; treat its output as a strong draft, not ground truth.
+
+## License & attribution
+
+Released under **Apache 2.0**, inheriting the base model's license. Built by **Leslie Moody**. Yi-Coder is by 01.AI. The coding-instruction data is from the dataset linked above.