Buckets:
| # Cursor Handoff: Mythos-Coder Dataset Project | |
| ## Project Goal | |
| I am building a coding-agent training dataset for a system called Mythos-Coder. | |
| The goal is to train/improve a coding LLM or coding agent for: | |
| - website building | |
| - vibe coding | |
| - repo editing | |
| - bug fixing | |
| - terminal debugging | |
| - feature building | |
| - refactoring | |
| - migrations | |
| - test writing | |
| - browser verification | |
| - long-horizon coding tasks | |
| This is not an AGI project and not a cybersecurity project. | |
| ## Current Dataset Format | |
| The main training dataset uses a canonical JSONL schema. | |
| Existing canonical fields include: | |
| - id | |
| - task_type | |
| - difficulty | |
| - language | |
| - framework | |
| - user_prompt | |
| - repo_context | |
| - failure_log | |
| - investigation_steps | |
| - plan | |
| - solution | |
| - verification | |
| - lesson | |
| - quality_score | |
| Raw idea rows should not be directly appended to the main training file unless converted first. | |
| ## Current Converter | |
| There is a Python converter script: | |
| convert_raw_ideas.py | |
| It converts raw idea JSONL rows from data/raw_ideas/ into the canonical mythos_coder_train.jsonl schema. | |
| The converter should be used before merging raw rows into the final dataset. | |
| ## Current Workflow | |
| 1. Generate raw idea rows from public repos. | |
| 2. Save them into data/raw_ideas/. | |
| 3. Validate the raw JSONL. | |
| 4. Convert them using convert_raw_ideas.py. | |
| 5. Write converted rows to a standalone output first. | |
| 6. Review the converted rows. | |
| 7. Only then append to the main dataset. | |
| ## Important Rule | |
| Do not generate huge batches all at once. | |
| Use small batches: | |
| - 10 rows at a time | |
| - validate | |
| - convert | |
| - review | |
| - repeat | |
| This avoids memory problems and bad rows. | |
| ## Repos Already Discussed | |
| ### 1. bradtraversy/50projects50days | |
| Best for: | |
| - feature_build | |
| - bug_fix | |
| - ui_polish | |
| - browser_verification | |
| - multi_file_edit | |
| ### 2. zce/html5up | |
| Best for: | |
| - website_vibe | |
| - ui_polish | |
| - browser_verification | |
| - multi_file_edit | |
| - responsive layout | |
| - landing page polish | |
| ## Next Target Repo | |
| Use: | |
| https://github.com/zce/html5up | |
| Create raw rows in: | |
| data/raw_ideas/html5up_website_vibe_rows.jsonl | |
| Convert to standalone: | |
| datasets/html5up_website_vibe_converted.jsonl | |
| Use id prefix: | |
| html5up-vibe | |
| ## Safety Rules | |
| - Do not copy source code from public repos. | |
| - Use repos only as inspiration. | |
| - Rewrite all rows in original words. | |
| - Keep raw rows separate from canonical rows. | |
| - Do not append to datasets/mythos_coder_train.jsonl until the standalone converted rows look valid. |
Xet Storage Details
- Size:
- 2.41 kB
- Xet hash:
- e4bfd3b66d91d538499f6aa7942e48a320c7a65419bc717aa170803cb1d21d44
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.