Spaces:
Sleeping
Sleeping
Commit ·
bdb156f
1
Parent(s): e5661c5
feat: initial structure
Browse files- Dockerfile +8 -0
- data/corruption.py +0 -0
- data/generator.py +0 -0
- data/loader.py +0 -0
- environment/__init__.py +0 -0
- environment/actions.py +0 -0
- environment/env.py +0 -0
- environment/reward.py +0 -0
- environment/server.py +0 -0
- eval/baseline_eval.py +0 -0
- openenv.yaml +35 -0
- training/train_grpo.py +0 -0
Dockerfile
ADDED
|
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
FROM python:3.11-slim
|
| 2 |
+
WORKDIR /app
|
| 3 |
+
COPY requirements.txt .
|
| 4 |
+
RUN pip install --no-cache-dir -r requirements.txt
|
| 5 |
+
COPY . .
|
| 6 |
+
RUN python -c "from data.loader import build_fact_database; build_fact_database()" || true
|
| 7 |
+
EXPOSE 7860
|
| 8 |
+
CMD ["uvicorn", "environment.server:app", "--host", "0.0.0.0", "--port", "7860"]
|
data/corruption.py
ADDED
|
File without changes
|
data/generator.py
ADDED
|
File without changes
|
data/loader.py
ADDED
|
File without changes
|
environment/__init__.py
ADDED
|
File without changes
|
environment/actions.py
ADDED
|
File without changes
|
environment/env.py
ADDED
|
File without changes
|
environment/reward.py
ADDED
|
File without changes
|
environment/server.py
ADDED
|
File without changes
|
eval/baseline_eval.py
ADDED
|
File without changes
|
openenv.yaml
ADDED
|
@@ -0,0 +1,35 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
name: context-corruption-env
|
| 2 |
+
version: "1.0.0"
|
| 3 |
+
description: >
|
| 4 |
+
OpenEnv environment for training epistemic robustness in LLMs.
|
| 5 |
+
Agents identify correct answers and flag corrupted documents
|
| 6 |
+
in a multi-doc QA setting with verifiable, objective rewards.
|
| 7 |
+
author: "Your Team Name"
|
| 8 |
+
license: MIT
|
| 9 |
+
|
| 10 |
+
environment:
|
| 11 |
+
entrypoint: "environment.server:app"
|
| 12 |
+
action_schema: "environment.actions.ContextCorruptionAction"
|
| 13 |
+
observation_schema: "environment.actions.EpisodeObservation"
|
| 14 |
+
max_concurrent_sessions: 64
|
| 15 |
+
|
| 16 |
+
reward:
|
| 17 |
+
type: "objective"
|
| 18 |
+
range: [-0.5, 1.05]
|
| 19 |
+
components:
|
| 20 |
+
- {name: answer_correctness, weight: 0.4}
|
| 21 |
+
- {name: corruption_detection, weight: 0.3}
|
| 22 |
+
- {name: false_positive_penalty, weight: 0.2}
|
| 23 |
+
- {name: confidence_calibration, weight: 0.1}
|
| 24 |
+
|
| 25 |
+
datasets:
|
| 26 |
+
- {name: Natural Questions, url: "https://huggingface.co/datasets/google-research-datasets/natural_questions"}
|
| 27 |
+
- {name: PopQA, url: "https://huggingface.co/datasets/akariasai/PopQA"}
|
| 28 |
+
- {name: FaithEval, url: "https://github.com/SalesforceAIResearch/FaithEval"}
|
| 29 |
+
|
| 30 |
+
citation: |
|
| 31 |
+
@misc{contextcorruption2026,
|
| 32 |
+
title={ContextCorruption-Env: Training Epistemic Robustness in LLMs},
|
| 33 |
+
year={2026},
|
| 34 |
+
note={OpenEnv Hackathon Submission}
|
| 35 |
+
}
|
training/train_grpo.py
ADDED
|
File without changes
|