File size: 2,605 Bytes
ddbc1ba | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 | # Scripts reference
**Directory:** `scripts/`
---
## `scripts/train_trl.py` — main training
Full GRPO training — single-step curriculum and episodic curriculum. See [train_trl.md](train_trl.md) for complete reference.
```bash
LIFESTACK_NO_UNSLOTH=1 python scripts/train_trl.py # full 5-stage curriculum
python scripts/train_trl.py --dry-run # pipeline check, CPU OK
python scripts/train_trl.py --resume # resume from checkpoint
python scripts/train_trl.py --episode-train # episodic mode
python scripts/train_trl.py --push-to-hub --hub-model-id ... # upload after training
```
---
## `scripts/eval.py` — random baseline
Runs a uniform random policy for N episodes and reports mean/std reward. No trained model, no GPU, no API key needed. Use this to establish a reward floor before GRPO runs or to verify env correctness after code changes.
```bash
python scripts/eval.py # 10 episodes, all domains
python scripts/eval.py --episodes 20 --domain flight_crisis
python scripts/eval.py --episodes 5 --verbose # per-step output
```
---
## `scripts/plot_training.py` — plot generation
Parses training logs (`train_run_v*.log`) and generates matplotlib plots: reward curve, loss curve, per-component reward breakdown, and 4-panel summary. Supports multiple log formats.
```bash
python scripts/plot_training.py --log train_run_v4.log --output-dir plots/
```
---
## `scripts/smoke_test.py`
Fast pipeline check: import validation + one `reset()` + one `step()`. No GPU, no downloads.
```bash
python scripts/smoke_test.py
```
Also called by `setup.sh` at the end of the install process.
---
## `scripts/upload_hf_model_cards.py`
Uploads model cards and training artifacts to HuggingFace model repositories:
- `docs/HF_MODEL_CARD_V4.md` → `jdsb06/lifestack-grpo-v4/README.md`
- `docs/HF_MODEL_CARD_V1.md` → `jdsb06/lifestack-grpo/README.md`
- `train_run_v1.log` → `jdsb06/lifestack-grpo/`
- Relevant plots → `jdsb06/lifestack-grpo-v4/plots/`
```bash
python scripts/upload_hf_model_cards.py # requires HF_TOKEN or huggingface-cli login
```
---
## `scripts/run_episode.py`
Runs one full episode with the `LifeStackAgent` (GRPO model or Groq API fallback). Requires credentials configured.
```bash
python scripts/run_episode.py
python scripts/run_episode.py --difficulty 3 --verbose
```
---
## Related files
- `docs/train_trl.md` — full train_trl.py reference
- `docs/eval.md` — eval.py reference
- `docs/training_guide.md` — end-to-end training guide
|