| --- |
| license: apache-2.0 |
| tags: |
| - minecraft |
| - rlcd |
| - decision-making |
| - fine-tuning-data |
| --- |
| |
| # Minecraft System-1 decision logs — Laya fine-tuning corpora |
|
|
| Decision ticks recorded by a Minecraft agent whose System-1 brain is [Laya](https://huggingface.co/convaiinnovations/laya). Each line is one tick of the agent loop (`perceive → decide → execute → verify`): a compact English narrative of the world state, the typed questions asked of the model, and the label the verifier accepted. |
|
|
| **Reward semantics:** only ticks whose action demonstrably made progress become positive examples (pure imitation with `--include-failed` is possible in the builder). Ping-pong ticks (six or more ticks repeating one label with no item gained) teach no intent and are filtered out. |
|
|
| ## Schema (JSONL, one object per line) |
|
|
| ```json |
| { |
| "state": "You are a Minecraft player ... Inventory: 4x acacia_log ...", |
| "questions": { |
| "intent": {"type": "choice", "instructions": "...", "criteria": {"craft_pickaxe": "...", "craft_table": "..."}}, |
| "in_danger": {"type": "choice", "instructions": "...", "criteria": {"no": "...", "yes": "..."}} |
| }, |
| "expected": {"intent": "craft_pickaxe", "in_danger": "no"} |
| } |
| ``` |
|
|
| ## Files |
|
|
| | File | Items | What it is | |
| |---|---|---| |
| | `train.jsonl` | 4,631 | Main corpus (runs up to Oct 2026) | |
| | `eval.jsonl` | 12 | Small held-out eval | |
| | `ood-eval.jsonl` | 33 | Out-of-distribution eval (endgame states the base model has never seen) | |
| | `run6.jsonl`, `run6-eval.jsonl`, `run6-preview.jsonl` | 56 / 18 / 56 | Earliest smoke dataset (run 6) | |
| | `run79-86-train.jsonl`, `run79-86-eval.jsonl` | 512 / 171 | M3 second round (runs 79–86) | |
| | `train-187-194.jsonl`, `eval-187-194.jsonl` | 609 / 203 | Runs 187–194 | |
| | `train-187-205.jsonl`, `eval-206-208.jsonl` | 1,509 / 219 | Runs 187–205 (basis of the 2026-11-07 checkpoint) | |
| | `relative-train.jsonl`, `relative-eval.jsonl` | 849 / 283 | Relative-narrative experiment (M5) | |
| | `train-smoke.jsonl`, `train-test.jsonl` | 38 / 116 | Smoke/test sets for the training pipeline | |
|
|
| ## How the data was built |
|
|
| Run logs (`decisions.jsonl` per run) are turned into these corpora with the agent's dataset builder: |
|
|
| ```bash |
| .venv/bin/python -m mlaya.cli dataset --log runs/history/decisions-run79.jsonl \ |
| --log runs/history/decisions-run80.jsonl --out train.jsonl --eval eval.jsonl |
| ``` |
|
|
| Agent code: [github.com/thtesche/minecraft-laya-training](https://github.com/thtesche/minecraft-laya-training) · Fine-tuned weights: [thtesche/laya-minecraft-english](https://huggingface.co/thtesche/laya-minecraft-english) |
|
|
