---
pretty_name: "train_rl completion logs"
tags:
- train_rl
- completions
- on-policy
- grpo
---

# train_rl Completion Logs

This dataset contains the on-policy generations produced during RL training
with [`train_rl`](https://github.com/kashif/train_rl).

## Training details

| Key | Value |
|-----|-------|
| **Algorithm** | GRPO |
| **Model (student)** | [Qwen/Qwen3-4B-Instruct-2507](https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507) |
| **Prompt dataset** | VerifierEnvDataset |
| **Group size** | 4 |
| **Max completion tokens** | 512 |
| **Temperature** | 1.0 |
| **Learning rate** | 1e-05 |

## Schema

Each parquet file corresponds to one rollout step and contains the following
columns:

| Column | Type | Description |
|--------|------|-------------|
| `step` | int | Rollout step index |
| `prompt` | str | Input prompt text |
| `completion` | str | Model-generated completion |
| `reward` | float | Reward for this completion |
| `advantage` | float / list[float] | Computed advantage (scalar for GRPO, per-token for OPD) |
| `response_length` | int | Number of response tokens |

## Loading the data

```python
from datasets import load_dataset

ds = load_dataset("kashif/train_rl_agent_completions")
```

```python
import polars as pl

df = pl.read_parquet("hf://datasets/kashif/train_rl_agent_completions/*.parquet")
```