README.md

metadata

license: apache-2.0
tags:
  - agents
  - tool-use
  - trajectories
  - nethack
  - minihack
configs:
  - config_name: default
    data_files:
      - split: train
        path: episodes/**/*.json

Lantern episodes

Every episode played in Lantern: Beam (and later a distilled 1.7B student) exploring MiniHack dungeon levels through tool calls, with its full reasoning, plans, notes and lessons.

Layout

episodes/{level}/{id}.json, one file per episode, plus journal.json (the Space's current lessons per level).

Each episode has id, task (level), seed, outcome (success, failure, timeout, budget, stalled, stopped, error), steps, use_lessons (journal on/off), source, model, and events:

• start: goal, legal actions, the exact system prompt and tool schema, model, prompt version, and the journal lessons read before the run.
• turn (one per model call): the observation prompt, the model's reasoning, the tool calls it made (planned moves, moves actually done, whether the plan was cut short, the hero's position after every move), its notes, the resulting map, running token usage, and any advice a viewer sent.
• lessons: the lessons written after the run.
• end: outcome, steps, token usage.

Sources

• bench: the benchmark (tag bench-v2): 8 levels × 5 seeds × journal off/on, all with Beam and prompt version 2. These 80 episodes are the main data.
• space: live runs started by visitors to the Space. coached: true marks runs where the visitor sent advice, and that advice text is included.
• student-eval: the distilled student and the untrained base model on the benchmark seeds.

Episodes with outcome error (infrastructure failures, replayed afterwards) or stopped aren't results; leave them out when computing win rates.

Using it for training

To build supervised training rows, take winning episodes made with the current prompt version and skip coached runs. Each turn becomes one example: the system prompt from start plus that turn's observation as input, and the tool calls the model made as output. This is how Lantern-1.7B was trained.

Privacy

Visitors' IP addresses are never stored. Space episodes contain only the run itself and any advice text a visitor chose to send.

Limitations

• Small: 80 benchmark episodes, 5 seeds per level and condition.
• Beam's reasoning is included as returned by the provider; it isn't edited.
• Level goals include some strategy hints (see the Space README).

License

Apache-2.0. Episodes contain outputs generated by Beam 1.0 (Reflection, Apache-2.0) and Lantern-1.7B, on levels from MiniHack (Apache-2.0).