Update ProRL usage notes
Browse files
README.md
CHANGED
|
@@ -107,6 +107,42 @@ In Harness-RL, the no-Docker launcher can pass held-out data to Slime eval with
|
|
| 107 |
PROMPT_DATA=data/train.jsonl PROMPT_MANIFEST=data/train_manifest.jsonl EVAL_PROMPT_DATA="heldout data/eval.jsonl" EVAL_INTERVAL=5 N_SAMPLES_PER_EVAL_PROMPT=1 EVAL_MAX_RESPONSE_LEN=4096 bash examples/harnessmix/no_docker_smoke/run_train.sh
|
| 108 |
```
|
| 109 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 110 |
## Provenance and Licensing
|
| 111 |
|
| 112 |
This dataset is a processed mix of multiple upstream public datasets and benchmarks. Users are responsible for complying with each upstream dataset's license and usage terms. This repository does not claim a single unified license over upstream content.
|
|
|
|
| 107 |
PROMPT_DATA=data/train.jsonl PROMPT_MANIFEST=data/train_manifest.jsonl EVAL_PROMPT_DATA="heldout data/eval.jsonl" EVAL_INTERVAL=5 N_SAMPLES_PER_EVAL_PROMPT=1 EVAL_MAX_RESPONSE_LEN=4096 bash examples/harnessmix/no_docker_smoke/run_train.sh
|
| 108 |
```
|
| 109 |
|
| 110 |
+
## ProRL-Agent-Server / Polar Slime Usage
|
| 111 |
+
|
| 112 |
+
The ProRL `codegen` branch should track code, configs, and documentation only. It should not duplicate this dataset in Git. From a ProRL-Agent-Server checkout, download the dataset into the codegen example directory before running smoke tests or GRPO training:
|
| 113 |
+
|
| 114 |
+
```bash
|
| 115 |
+
export HF_TOKEN=<your-huggingface-token> # only needed if this dataset is private
|
| 116 |
+
python - <<'PY_INNER'
|
| 117 |
+
import os
|
| 118 |
+
from huggingface_hub import snapshot_download
|
| 119 |
+
|
| 120 |
+
snapshot_download(
|
| 121 |
+
repo_id="Xnhyacinth/codegen",
|
| 122 |
+
repo_type="dataset",
|
| 123 |
+
local_dir="examples/codegen_slime_grpo",
|
| 124 |
+
token=os.environ.get("HF_TOKEN"),
|
| 125 |
+
allow_patterns=["data/**", "assets/**", "metadata/**", "raw/**"],
|
| 126 |
+
)
|
| 127 |
+
PY_INNER
|
| 128 |
+
```
|
| 129 |
+
|
| 130 |
+
After download, the ProRL codegen example expects these local paths:
|
| 131 |
+
|
| 132 |
+
- `examples/codegen_slime_grpo/data/train.jsonl`
|
| 133 |
+
- `examples/codegen_slime_grpo/data/eval.jsonl`
|
| 134 |
+
- `examples/codegen_slime_grpo/assets/`
|
| 135 |
+
- `examples/codegen_slime_grpo/metadata/`
|
| 136 |
+
- `examples/codegen_slime_grpo/raw/ojbench_testdata/` for the current OJBench subset
|
| 137 |
+
|
| 138 |
+
Minimal ProRL smoke command:
|
| 139 |
+
|
| 140 |
+
```bash
|
| 141 |
+
RUN_ID=codegen-smoke PROMPT_DATA=examples/codegen_slime_grpo/data/train_smoke.jsonl REF_LOAD=/path/to/Qwen3.5-4B_torch_dist WANDB_MODE=offline CUDA_VISIBLE_DEVICES=0,1,2,3 bash examples/codegen_slime_grpo/run.sh
|
| 142 |
+
```
|
| 143 |
+
|
| 144 |
+
For full training, use `PROMPT_DATA=data/train.jsonl` and `EVAL_PROMPT_DATA=data/eval.jsonl` through the ProRL codegen launcher. Keep W&B logs, checkpoints, Ray temp files, and downloaded dataset files outside Git.
|
| 145 |
+
|
| 146 |
## Provenance and Licensing
|
| 147 |
|
| 148 |
This dataset is a processed mix of multiple upstream public datasets and benchmarks. Users are responsible for complying with each upstream dataset's license and usage terms. This repository does not claim a single unified license over upstream content.
|