README.md

---
license: apache-2.0
language:
- en
task_categories:
- text-generation
tags:
- math
- reinforcement-learning
- tool-use
- hindsight
configs:
- config_name: source
  data_files:
  - split: full
    path: rl_trace_opsd.parquet
- config_name: grpo
  default: true
  data_files:
  - split: train
    path: grpo/train.parquet
  - split: validation
    path: grpo/validation.parquet
---

# rl_trace_opsd

The versioned mathematical RL data used by the ToolElite TRACE-OPSD project.
The source contains **20,161 examples**, with columns `id`, `problem`,
`ground_truth`, `difficulty`, `hindsight`, and `solution`. Difficulty labels are
`mid` (10,940 examples) and `high` (9,221 examples).

## Two representations of the same examples

| Configuration | Split | Examples | Purpose |
|---|---|---:|---|
| `source` | `full` | 20,161 | Full problem pool, including held-out validation IDs |
| `grpo` | `train` | 19,905 | Immutable training split in ToolElite/verl format |
| `grpo` | `validation` | 256 | Fixed training-monitoring validation split |

These are different representations of the same pool, not additional datasets.
Do not train on the `source/full` split without excluding the fixed validation
IDs. The validation split has **160 mid + 96 high** examples; split seed is 42.
It is separate from final GSM8K/MATH500/AIME benchmark evaluation.

```python
from datasets import load_dataset

training = load_dataset("BitZone910/rl_trace_opsd", "grpo", split="train")
validation = load_dataset("BitZone910/rl_trace_opsd", "grpo", split="validation")
source = load_dataset("BitZone910/rl_trace_opsd", "source", split="full")
```

The GRPO format includes the tool prompt, terminal answer reward, source IDs,
and training-only privileged hindsight. At inference/evaluation, expose only the
problem and tool prompt to the policy; do not leak answer or hindsight fields.

## Source and attribution

Derived from [NuminaMath-TIR](https://huggingface.co/datasets/AI-MO/NuminaMath-TIR),
whose publisher declares Apache-2.0. ToolElite adds difficulty labels, solution
hindsight, filtering, and tool-training formatting. Preserve upstream attribution
and the included Apache license when redistributing.