Upload model README

README.md

@@ -1,3 +1,128 @@
----
-license: mit
----
+## sm_llm Workflow
+
+This folder contains the end-to-end pipeline to fine-tune `Qwen/Qwen2.5-Coder-1.5B-Instruct`
+on SuperMongo examples and export a quantized GGUF model for Ollama.
+
+### Files
+
+- `1_dataset.jsonl`: ChatML training set
+- `1b_alignment.jsonl`: targeted alignment examples (identity + anti-pseudo-syntax)
+- `2_finetune_unsloth.py`: LoRA fine-tuning (Unsloth)
+- `3_merge_weights.py`: merge LoRA into full model
+- `4_convert_gguf.py`: HF -> GGUF + Q4_K_M quantization
+- `5_build_mlc.py`: build MLC/WebLLM artifacts from merged model
+- `6_publish_hf.py`: publish GGUF and/or MLC artifacts to Hugging Face
+- `Modelfile.finetuned`: Ollama model definition
+
+### Prerequisites (WSL + NVIDIA)
+
+Run training from Linux/WSL with CUDA available.
+
+```bash
+nvidia-smi
+python3 -c "import torch; print(torch.cuda.is_available())"
+```
+
+### Environment Setup
+
+From `sm_llm/`:
+
+```bash
+uv sync --extra train
+```
+
+If you are using an already active venv, use `--active` in `uv run`.
+
+### Build llama.cpp (submodule)
+
+The conversion script expects `llama.cpp` at `./llama.cpp` (submodule path).
+
+```bash
+cd llama.cpp
+cmake -B build
+cmake --build build -j
+cd ..
+```
+
+### Train + Export
+
+```bash
+uv run --active --extra train python 2_finetune_unsloth.py
+uv run --active --extra train python 3_merge_weights.py
+uv run --active --extra train python 4_convert_gguf.py
+```
+
+Generated artifacts:
+
+- `lora_model/`
+- `merged_model/`
+- `gguf/sm-coder-1.5b-q4_k_m.gguf`
+
+Note: `2_finetune_unsloth.py` loads both `1_dataset.jsonl` and `1b_alignment.jsonl` if present.
+
+### Ollama Test
+
+```bash
+ollama create sm-coder -f Modelfile.finetuned
+ollama run sm-coder
+```
+
+### Build MLC (WebLLM)
+
+Install MLC CLI (inside WSL, in your active env).
+The packages are hosted on a custom index and require platform-specific names.
+For CUDA 13.0 (RTX 4080 / sm89):
+
+```bash
+uv pip install --pre -U --find-links https://mlc.ai/wheels mlc-llm-nightly-cu130 mlc-ai-nightly-cu130
+```
+
+> Note: `uv`-created venvs do not include `pip` by default, so use `uv pip` instead of `python -m pip`.
+> For other CUDA versions replace `cu130` with e.g. `cu128` or `cpu`.
+
+Build MLC artifacts from `merged_model/`:
+
+```bash
+uv run --active python 5_build_mlc.py --model-id your-user/sm-coder-1.5b
+```
+
+Output is written to:
+
+- `mlc_dist/weights`
+- `mlc_dist/lib/*.wasm`
+
+### Publish on Hugging Face
+
+Install publish dependencies:
+
+```bash
+uv sync --extra publish
+```
+
+Login once:
+
+```bash
+huggingface-cli login
+```
+
+Publish only GGUF (the model used by Ollama):
+
+```bash
+uv run --active --extra publish python 6_publish_hf.py \
+	--gguf-repo your-user/sm-coder-gguf
+```
+
+Publish both GGUF and MLC artifacts:
+
+```bash
+uv run --active --extra publish python 6_publish_hf.py \
+	--gguf-repo your-user/sm-coder-gguf \
+	--mlc-repo your-user/sm-coder-mlc
+```
+
+Use `--private` if you want private repositories.
+
+### Notes
+
+- On `/mnt/c` you may see UV hardlink warnings; they are harmless.
+- You can silence them with: `export UV_LINK_MODE=copy`