---
language:
- en
tags:
- evaluation
- reasoning
- llm
- science
- mathematics
size_categories:
- 10K<n<100K
configs:
- config_name: hle
  data_files:
  - split: test
    path: data/hle/test-*.parquet
---

# Lite-RSI

Lite-RSI is a multi-benchmark collection of normalized model responses and automated evaluation traces. The current release provides the `hle` configuration. Each row is one model response to one task. Repeated `task_id` values are intentional: they represent responses from different models or generation runs.

## Available Configurations

| Config | Split | Description |
| --- | --- | --- |
| `hle` | `test` | HLE multi-model evaluation responses. |

## Data Format

The dataset is published as a single standard Hugging Face Parquet shard:

```
data/hle/test-00000-of-00001.parquet
```

`record_id` is a unique release-level identifier. `task_id` identifies the underlying question. `model_name` identifies the normalized generating model.

| Field | Description |
| --- | --- |
| `record_id` | Unique normalized record identifier. |
| `task_id` | Underlying task identifier; repeats across model responses. |
| `model_name` | Generating model, with provider-prefix aliases normalized. |
| `generation_source` | Original/retry/gold/run-source label. |
| `prompt` | User prompt text. |
| `reference_answer` | Reference target from the original task. |
| `response` | Model completion. |
| `judge_label`, `judge_confidence`, `judge_answer`, `judge_explanation` | Existing automated-grader fields. |
| `subject`, `category`, `answer_type`, `has_image`, `rationale` | Original task metadata. |
| `sample_sha256` | Integrity hash of the original sample. |

## Loading

```python
from datasets import load_dataset

dataset = load_dataset(
    "lhpku20010120/Lite-RSI",
    "hle",
    split="test",
)
```

## Data Notes

- Empty completions are retained and marked by `has_response=false`.
- This release does not deduplicate model outputs; comparisons across models and retries are a primary use case.
- Additional benchmarks will be published as separate dataset configurations under `data/<benchmark>/`.
- Before publishing, the dataset owner must confirm the licenses and redistribution rights of the original questions, answers, model outputs, and any referenced attachments.