README.md

---
license: mit
task_categories:
- text-retrieval
language:
- en
size_categories:
- 1M<n<10M
dataset_info:
  features:
  - name: id
    dtype: string
  - name: url
    dtype: string
  - name: title
    dtype: string
  - name: text
    dtype: string
  - name: edu_quality
    dtype: float64
  - name: naive_quality
    dtype: int64
  splits:
  - name: train
    num_bytes: 1820347708
    num_examples: 1041047
  download_size: 1104502937
  dataset_size: 1820347708
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---

# MTR Document Collection

**1,041,047** Wikipedia passages used as the retrieval corpus for the MTR benchmark.

Part of [MTR-Suite](https://github.com/OkayestProgrammer/mtr-suite) (ACL 2026 Main).

## Related Resources

| Resource | Link | Description |
|----------|------|-------------|
| 📊 **MTR Benchmark** | [`OkayestProgrammer/MTR-BENCH`](https://huggingface.co/datasets/OkayestProgrammer/MTR-BENCH) | Original test set |
| 🏋️ **MTR Training** | [`OkayestProgrammer/MTR-train`](https://huggingface.co/datasets/OkayestProgrammer/MTR-train) | Original training set |
| 🆕 **12-Turn Dataset** | [`OkayestProgrammer/mtr-qwen35-fp8-12turn`](https://huggingface.co/datasets/OkayestProgrammer/mtr-qwen35-fp8-12turn) | 10K 12-turn conversations with topic switches |
| 🔧 **Code** | [`OkayestProgrammer/mtr-suite`](https://github.com/OkayestProgrammer/mtr-suite) | Full pipeline code |

## Columns

| Column | Type | Description |
|--------|------|-------------|
| `id` | string | Wikipedia article ID |
| `url` | string | Source URL |
| `title` | string | Article title |
| `text` | string | Passage text (max 2048 chars) |
| `edu_quality` | float | Educational quality score |
| `naive_quality` | int | Naive quality label |

## Usage

```python
from datasets import load_dataset

docs = load_dataset("OkayestProgrammer/MTR-DOCUMENT", split="train")
print(f"{len(docs)} documents")
print(docs[0]["title"], "-", docs[0]["text"][:100])
```

## How this corpus is used

The `ground_truth_document_idx` field in the MTR query datasets (e.g., `MTR-BENCH`, `mtr-qwen35-fp8-12turn`) is a **row index** into this document collection. During evaluation:

1. Embed all 1,041,047 documents → build FAISS index
2. Embed test queries → search the index
3. Compare retrieved document indices against `ground_truth_document_idx`

See the [eval guide](https://huggingface.co/datasets/OkayestProgrammer/mtr-qwen35-fp8-12turn#evaluation-with-mtr-suite) for step-by-step instructions.

## Citation

```bibtex
@inproceedings{mtr-suite-2026,
    title={MTR-Suite: A Data Synthesis Pipeline, Benchmark, and Models for Conversational Retrieval},
    author={},
    booktitle={Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026)},
    year={2026}
}
```