---
license: mit
tags:
- openenv
- reinforcement-learning
- grpo
- unsloth
- evacuation
---

# EvacOS2 Public Training Artifacts

This repository hosts the public LoRA adapter artifacts, metrics, logs, and generated configs selected for the EvacOS2 OpenEnv India Hackathon 2026 submission.

Canonical environment Space: https://huggingface.co/spaces/shashankN777/evacos2-openenv

Canonical code repo: https://github.com/sai-shashankN/EvacOS2

## Evidence Status

The visible submitted checkpoints in this repo are `3B` floor-specialist H200 canaries for fire, flood, and gas response lanes. They are not claimed as final `7B` orchestrator quality checkpoints. The tracked fixed-suite scorecard in the code repo is baseline-only; the tracked learning evidence is the fire/flood/gas specialist canary and training-signal artifact trail.

## Public Checkpoint Paths

| Specialist | Public checkpoint path | Run type |
|---|---|---|
| Fire floor specialist | `floor-specialists/fire/h200-canary3-10/checkpoints/latest` | `10`-step H200 canary |
| Flood floor specialist | `floor-specialists/flood/h200-canary3-10/checkpoints/latest` | `10`-step H200 canary |
| Gas floor specialist | `floor-specialists/gas/h200-canary-10/checkpoints/latest` | `10`-step H200 canary |

Each run folder includes:

- `checkpoints/latest/lora_adapter/floor_agent/adapter_model.safetensors`
- `checkpoints/latest/lora_adapter/floor_agent/adapter_config.json`
- generated training config YAML
- training metrics CSV
- JSONL traces and summaries
- train/contrast exit-code markers

## Download

```bash
hf download shashankN777/evacos2-7b-orchestrator-artifacts \
  --include "floor-specialists/fire/h200-canary3-10/**" \
  --include "floor-specialists/flood/h200-canary3-10/**" \
  --include "floor-specialists/gas/h200-canary-10/**" \
  --local-dir outputs/hf_public_artifacts
```

## Evaluate A Restored Canary Checkpoint

Run this from the EvacOS2 code repo after downloading the artifacts:

```bash
CHECKPOINT_DIR=outputs/hf_public_artifacts/floor-specialists/fire/h200-canary3-10/checkpoints/latest
CONFIG_PATH=outputs/hf_public_artifacts/floor-specialists/fire/h200-canary3-10/generated.remote-unsloth-3b-fire-floor-specialist-h200-canary3-10.yaml
METRICS_PATH=outputs/hf_public_artifacts/floor-specialists/fire/h200-canary3-10/remote-unsloth-3b-fire-floor-specialist-h200-canary3-10-metrics.csv

python -m evaluation.demo_bundle \
  --trained-checkpoint "$CHECKPOINT_DIR" \
  --config "$CONFIG_PATH" \
  --training-metrics-path "$METRICS_PATH" \
  --output-dir outputs/demo_bundle_fire_h200_canary
```

For the checked-in baseline-only fixed suite, see `demo/results/submission_scorecard_baseline.md` in the code repo. For the tracked learning-signal summary, see `demo/results/specialist_canary50_report.md` and `demo/results/3b_specialist_canary50_scores.csv`.