README.md

---
license: other
language: [pl]
pretty_name: Warsaw Model Trainers Hackathon data (history matura)
---
# hackathon-data

Datasets and evaluation results of the Warsaw Model Trainers Hackathon (Polish history matura), moved out of the
GitHub repo `Endote/WarsawModelTrainersHackathon`. Restore with `scripts/fetch_data.py` in that repo.

Paths mirror the repo, so restored files land where the code expects them:

| Path | Content |
|---|---|
| `data/evaluation-2017-2026`, `data/evaluation-2005-2016`, `data/grading-2017-2026`, `data/arkusze` | exams, answer keys, grading packets |
| `data/epodreczniki`, `data/synthetic/rephrased` | source corpus and rephrased questions |
| `output/EVAL`, `output/feedback`, `output/assets`, `output/rag-dense`, `output/essay-evidence`, `output/history-v2-*` | evaluation runs, feedback, RAG and training-run artifacts |
| `tmp/*` | grading and source-graph working sets |
| `output/essay-generation/history-500-v1..v8` | essay generation runs (packets, attempts, gradings). `v7` is the versioned source of the completed `WMTH-100d2exam/synthetic-history-essays` release; the other runs are partial. **Never ingest partial runs, and never use `key_answer` as an assistant target** (see the note below). |
| `data_synthetic_extended/` | the extended synthetic essays and questions (`generated/`, `grader_calibration/`, `inventory/`, docs and scripts). **Deprecated as a training source**, kept for reference only: `generated/*/essay_*.jsonl` holds examiner guidance in `key_answer`. |
| `synthetic-history-essays/` | the completed 500-essay release (`history-500-v7`, essays scored at least 10/15, plus solver/grader/practice configs), including the `provenance` config (its `original_json` embeds the original `data_synthetic_extended` records with examiner `key_answer` text: never use it as an assistant target). Rights: see its `RIGHTS.md`, no blanket license is asserted. |
| `hf-sources/<name>/` | the standalone HF datasets merged in (files identical to the ones above were dropped) |

Cleaning: server IPs, home paths and the ClearML API key were redacted (`X.X.X.X`, `/home/user`, `REDACTED_CLEARML_*`).
Deprecated sources are NOT included: `data_synthetic_image*` and `history-lora-data-2201`, and the
candidate catalog `synthetic-history-essays-v1` (it embeds examiner `key_answer` text). The completed release is included in full.

**Note on `data_synthetic_extended` and `output/essay-generation`:** the packets were built from the deprecated `data_synthetic_extended` essays, so many
files (about 7,500 in the runs, and the extended essays themselves) contain examiner guidance in `key_answer`. Both are published as audit evidence (the runs are also the source of
the completed essay release, which is not deprecated). `data_synthetic_extended` itself stays excluded from training. That guidance must never become an assistant essay target, and only
completed, reviewed answers from a versioned release may be used for training.
Scores on these exams are development diagnostics, not a held-out estimate.