chore: add HF Space frontmatter and README draft

README.md

@@ -1,10 +1,93 @@
 ---
 title: Context Corruption Env
-emoji: 📊
-colorFrom: gray
-colorTo: blue
+emoji: 🔍
+colorFrom: blue
+colorTo: purple
 sdk: docker
+app_port: 7860
 pinned: false
+license: mit
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# ContextCorruption-Env
+
+> OpenEnv Hackathon | Meta × HuggingFace × PyTorch
+
+An RL environment for training **epistemic robustness** in LLMs — teaching models to identify corrupted documents and resist misleading retrieved context.
+
+## The Problem
+
+LLMs fine-tuned with RLHF increasingly defer to retrieved documents even when those documents contradict the model's own (correct) parametric knowledge. Standard retrieval-augmented generation makes this worse. We train a model to cross-reference claims, flag corrupted sources, and answer from reliable knowledge.
+
+## The Environment
+
+- **8 documents** per episode, 1–4 deliberately corrupted
+- **4 corruption levels**: number mutation → entity swap → semantic inversion → coherent fabrication
+- **450 QA facts** from Natural Questions + PopQA
+- **Deterministic reward**, no LLM judge
+
+### Reward Signal
+
+| Component | Logic | Weight |
+|---|---|---|
+| Answer correctness | Exact match after normalisation | +0.40 |
+| Corruption detection recall | Fraction of corrupt docs flagged | +0.30 |
+| Precision (no false flags) | −0.1 per false positive, floor 0 | +0.20 |
+| Confidence calibration | +0.1×conf if correct, −0.2×conf if wrong | ±0.10 |
+| Efficiency bonus | Budget not wasted | +0.05 |
+
+**Range:** −0.5 to 1.05 · **Random baseline:** avg 0.1302
+
+## Results
+
+| Agent | Avg Reward | Answer Acc | Corruption Detection |
+|---|---|---|---|
+| Random baseline | 0.1302 | ~0% | ~25% |
+| Qwen2-1.5B GRPO | _after training_ | — | — |
+
+## Training
+
+```bash
+# Colab (A100 recommended)
+# Open training/ContextCorruption_GRPO.ipynb
+```
+
+Training script: [`training/train_grpo.py`](training/train_grpo.py)
+Notebook: [`training/ContextCorruption_GRPO.ipynb`](training/ContextCorruption_GRPO.ipynb)
+
+## Links
+
+- **HF Space (env server):** https://huggingface.co/spaces/Siddh12334/context-corruption-env
+- **Colab notebook:** _add link after training_
+- **WandB run:** _add link after training_
+- **Blog post:** _add link_
+
+## Quick Start
+
+```python
+from openenv.core import SyncEnvClient
+
+client = SyncEnvClient(base_url="https://Siddh12334-context-corruption-env.hf.space")
+obs = client.reset()
+print(obs["question"])
+```
+
+## Repo Structure
+
+```
+environment/   # OpenEnv-compliant env (actions, reward, env, server)
+data/          # QA loader, corruption functions, document generator
+training/      # GRPO training script + Colab notebook
+eval/          # Baseline evaluation
+assets/        # reward_curve.png, loss_curve.png (after training)
+```
+
+## Citation
+
+```bibtex
+@misc{contextcorruption2026,
+  title={ContextCorruption-Env: Training Epistemic Robustness in LLMs},
+  year={2026},
+  note={OpenEnv Hackathon Submission}
+}
+```