File size: 2,876 Bytes
86d386d | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 | ---
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.
|