Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,90 @@
|
|
| 1 |
-
---
|
| 2 |
-
license: cc-by-4.0
|
| 3 |
-
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: cc-by-4.0
|
| 3 |
+
---
|
| 4 |
+
---
|
| 5 |
+
pretty_name: DedeuceBench (Dev)
|
| 6 |
+
license: MIT
|
| 7 |
+
tags:
|
| 8 |
+
- llm-agents
|
| 9 |
+
- active-learning
|
| 10 |
+
- tool-use
|
| 11 |
+
- finite-state-machines
|
| 12 |
+
- benchmark
|
| 13 |
+
---
|
| 14 |
+
|
| 15 |
+
# DedeuceBench Dev Split
|
| 16 |
+
|
| 17 |
+
This dataset provides the public development split for DedeuceBench, an interactive active‑learning benchmark over hidden Mealy machines. Each item in the split is a config entry (seeded episode) defined in `levels_dev.json` under subset `dev`.
|
| 18 |
+
|
| 19 |
+
## Summary
|
| 20 |
+
- Task: Identification-first — agents must probe a hidden finite-state transducer under a strict query budget using tool calls (`act`, `submit_table`) and submit an exact transition table to succeed.
|
| 21 |
+
- Deterministic: Episodes are seeded; there is no ground-truth leak in the prompts.
|
| 22 |
+
- Split: `dev` subset mirrors an easy configuration (e.g., S=2, budget=25) for quick iteration and public sharing.
|
| 23 |
+
|
| 24 |
+
## Files
|
| 25 |
+
- `levels_dev.json`: multi-split JSON with a `dev` subset containing episode items by seed.
|
| 26 |
+
|
| 27 |
+
Example schema fragment:
|
| 28 |
+
```
|
| 29 |
+
{
|
| 30 |
+
"splits": {
|
| 31 |
+
"dev": {
|
| 32 |
+
"mode": "basic",
|
| 33 |
+
"trap": false,
|
| 34 |
+
"feedback": true,
|
| 35 |
+
"budget": 25,
|
| 36 |
+
"n_states": 2,
|
| 37 |
+
"target_len": 8,
|
| 38 |
+
"variety": false,
|
| 39 |
+
"items": [ {"seed": 1001}, ... ]
|
| 40 |
+
}
|
| 41 |
+
}
|
| 42 |
+
}
|
| 43 |
+
```
|
| 44 |
+
|
| 45 |
+
## Usage
|
| 46 |
+
Programmatic download:
|
| 47 |
+
```
|
| 48 |
+
pip install huggingface_hub
|
| 49 |
+
python - << 'PY'
|
| 50 |
+
from huggingface_hub import hf_hub_download
|
| 51 |
+
p = hf_hub_download(
|
| 52 |
+
repo_id="comfortably-dumb/dedeucebench-dev",
|
| 53 |
+
filename="levels_dev.json",
|
| 54 |
+
repo_type="dataset",
|
| 55 |
+
)
|
| 56 |
+
print(p)
|
| 57 |
+
PY
|
| 58 |
+
```
|
| 59 |
+
|
| 60 |
+
Evaluate with the DedeuceBench CLI (deterministic baseline):
|
| 61 |
+
```
|
| 62 |
+
dedeucebench-eval \
|
| 63 |
+
--split /path/to/levels_dev.json \
|
| 64 |
+
--subset dev \
|
| 65 |
+
--model heuristic:none \
|
| 66 |
+
--out results.dev.jsonl
|
| 67 |
+
|
| 68 |
+
dedeucebench-aggregate results.dev.jsonl > leaderboard.dev.csv
|
| 69 |
+
```
|
| 70 |
+
|
| 71 |
+
Using OpenRouter (OpenAI-compatible):
|
| 72 |
+
```
|
| 73 |
+
export OPENAI_API_KEY=... # your OpenRouter key
|
| 74 |
+
export OPENAI_BASE_URL=https://openrouter.ai/api/v1
|
| 75 |
+
|
| 76 |
+
dedeucebench-eval \
|
| 77 |
+
--split /path/to/levels_dev.json \
|
| 78 |
+
--subset dev \
|
| 79 |
+
--provider openrouter \
|
| 80 |
+
--model openai/gpt-5-mini \
|
| 81 |
+
--out results.openrouter_gpt5mini.dev.jsonl
|
| 82 |
+
```
|
| 83 |
+
|
| 84 |
+
## License
|
| 85 |
+
MIT
|
| 86 |
+
|
| 87 |
+
## Citation
|
| 88 |
+
- Concept DOI (all versions): 10.5281/zenodo.17166596
|
| 89 |
+
- See the main repo README for full citation details.
|
| 90 |
+
|