| # Extending MK-LLM | |
| This guide shows how to plug in different base models, datasets, and adapters. | |
| ## Swap base model | |
| - Set `MODEL_PATH` in `.env` to a local dir or HF repo id. | |
| - If using a HF repo, set `TRUST_REMOTE_CODE=true` when custom code is required. | |
| - Low-VRAM: set `LOAD_IN_4BIT=true` (or `LOAD_IN_8BIT=true`). | |
| ## Add datasets | |
| - Place cleaned text into `data/cleaned/*.txt` or generate `data/cleaned/mk_combined_data.txt` via `python -m data.process_all_data`. | |
| - The trainer uses `examples/data_loader.load_mk_dataset()` which prefers the combined file. | |
| ## Instruction tuning | |
| - Convert text into chat turns and use `tokenizer.apply_chat_template` in the training collator. | |
| - Provide Macedonian system prompts and stop sequences as needed. | |
| ## Custom inference params | |
| - Use `POST /v1/chat/completions` with `temperature`, `top_p`, `max_tokens`, `stream`. | |
| - Configure defaults via `.env`. | |
| ## Contribute plugins | |
| - Add new data collectors under `data/` and document flags in README. | |
| - Add new generation strategies or safety middlewares in `inference/`. | |