Dataset card
Browse files
README.md
ADDED
|
@@ -0,0 +1,56 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
tags:
|
| 4 |
+
- agents
|
| 5 |
+
- tool-use
|
| 6 |
+
- trajectories
|
| 7 |
+
- nethack
|
| 8 |
+
- minihack
|
| 9 |
+
pretty_name: Lantern episodes
|
| 10 |
+
---
|
| 11 |
+
|
| 12 |
+
# Lantern episodes
|
| 13 |
+
|
| 14 |
+
Every episode played in [Lantern](https://huggingface.co/spaces/visproj/lantern): Beam (and
|
| 15 |
+
later a distilled 1.7B student) exploring MiniHack dungeon levels through tool calls, with
|
| 16 |
+
its full reasoning, plans, notes and lessons.
|
| 17 |
+
|
| 18 |
+
## Layout
|
| 19 |
+
|
| 20 |
+
`episodes/{level}/{id}.json`, one file per episode, plus `journal.json` (the Space's current
|
| 21 |
+
lessons per level).
|
| 22 |
+
|
| 23 |
+
Each episode has `id`, `task` (level), `seed`, `outcome` (`success`, `failure`, `timeout`,
|
| 24 |
+
`budget`, `stalled`, `stopped`, `error`), `steps`, `use_lessons` (journal on/off), `source`,
|
| 25 |
+
`model`, and `events`:
|
| 26 |
+
|
| 27 |
+
- `start`: goal, legal actions, the exact system prompt and tool schema, model, prompt
|
| 28 |
+
version, and the journal lessons read before the run.
|
| 29 |
+
- `turn` (one per model call): the observation prompt, the model's reasoning, the tool calls
|
| 30 |
+
it made (planned moves, moves actually done, whether the plan was cut short, the hero's
|
| 31 |
+
position after every move), its notes, the resulting map, running token usage, and any
|
| 32 |
+
advice a viewer sent.
|
| 33 |
+
- `lessons`: the lessons written after the run.
|
| 34 |
+
- `end`: outcome, steps, token usage.
|
| 35 |
+
|
| 36 |
+
## Sources
|
| 37 |
+
|
| 38 |
+
- `bench`: the benchmark (tag `bench-v2`): 8 levels × 5 seeds × journal off/on, all with Beam
|
| 39 |
+
and prompt version 2. These 80 episodes are the main data.
|
| 40 |
+
- `space`: live runs started by visitors to the Space. `coached: true` marks runs where the
|
| 41 |
+
visitor sent advice.
|
| 42 |
+
- `student-eval`: the distilled student and the untrained base model on the benchmark seeds.
|
| 43 |
+
|
| 44 |
+
Episodes with `outcome` `error` (infrastructure failures, replayed afterwards) or `stopped`
|
| 45 |
+
aren't results; leave them out when computing win rates.
|
| 46 |
+
|
| 47 |
+
## Using it for training
|
| 48 |
+
|
| 49 |
+
`lantern/export.py` in the Space repo turns winning episodes into chat-format SFT rows (system
|
| 50 |
+
prompt and observation in, tool calls out). It skips coached runs and other prompt versions.
|
| 51 |
+
|
| 52 |
+
## Limitations
|
| 53 |
+
|
| 54 |
+
- Small: 80 benchmark episodes, 5 seeds per level and condition.
|
| 55 |
+
- Beam's reasoning is included as returned by the provider; it isn't edited.
|
| 56 |
+
- Level goals include some strategy hints (see the Space README).
|