| --- |
| license: cc-by-nc-nd-4.0 |
| language: |
| - en |
| tags: |
| - luna |
| - sft |
| - lora |
| - rag |
| - mcp |
| - training |
| --- |
| |
| # LUNA-Training — SFT & LoRA Recipes for LUNA-100M |
|
|
| This repository is the **supervised fine-tuning workspace** for the [LUNA-100M](https://huggingface.co/ASTERIZER/LUNA-100M) model. It contains the SFT/LoRA training code, configuration files, and the RAG + MCP dataset build pipeline used to produce the trained `sft_v1` / `luna_100m_sft` checkpoints. |
|
|
| ## Contents |
|
|
| | Path | Description | |
| |---|---| |
| | `train.py` | Base pretraining trainer (shared with the [LUNA workspace](https://huggingface.co/spaces/ASTERIZER/LUNA)) | |
| | `sft_train.py` | Full supervised fine-tuning trainer | |
| | `lora_sft_train.py` | LoRA SFT trainer (RAG+MCP adapter runs) | |
| | `lora_chat.py` / `chat.py` / `chat_full_sft.py` / `generate.py` | Inference / chat entry points | |
| | `validate_sft.py` / `validate_and_quantize.py` | Eval + quantization (incl. GGUF export) | |
| | `rag_mcp_lora_config.yaml` / `rag_mcp_full_sft_config.yaml` | RAG + MCP SFT recipes | |
| | `sft_config.yaml` / `train_config.yaml` | Base SFT / pretrain configs | |
| | `push_*_to_hf.py` | Upload helpers (code / dataset / model / LoRA) | |
| | `Base/Datasets/rag_mcp_sft/` | RAG + MCP SFT dataset builder + reports | |
| | `Base/checkpoints/EleutherAI/pythia-160m/` | LUNA tokenizer (vocab 50,254) | |
|
|
| ## Related |
|
|
| - **Model:** [ASTERIZER/LUNA-100M](https://huggingface.co/ASTERIZER/LUNA-100M) |
| - **SFT dataset:** [ASTERIZER/LUNA-RAG-MCP-SFT-10M](https://huggingface.co/datasets/ASTERIZER/LUNA-RAG-MCP-SFT-10M) |
| - **Pretrain corpus:** [ASTERIZER/LUNA_PreTrain](https://huggingface.co/datasets/ASTERIZER/LUNA_PreTrain) |
| - **Program collection:** [LUNA-100M Program](https://huggingface.co/collections/ASTERIZER/luna-100m-program-6a9115383c18e52460bf67c9) |
|
|
