---
pretty_name: Nuosu Task-Oriented Corpus
language:
- ii
- zh
- en
license: other
size_categories:
- 100K<n<1M
task_categories:
- text-generation
- translation
tags:
- nuosu
- sichuan-yi
- liangshan-yi
- standard-yi
configs:
- config_name: sft
  data_files: ready_sft.jsonl
- config_name: sft_validation
  data_files: validation_sft.jsonl
- config_name: cpt
  data_files: ready_cpt.jsonl
- config_name: research_test
  data_files: research_test_eval.jsonl
---

# Nuosu Task-Oriented Corpus / 诺苏语任务语料库

**Author:** Wuhe Axi, Xi'an Jiaotong University  
**Version:** 2026.08.04

## Fixed research split

This release uses a deterministic content-hash academic split derived from
`TianYeZ1214/NuosuBench@0709c1fd4f6eaabd4b058ea419027ccc42731dee`.

| File | Records | Gradient use |
|---|---:|---|
| `ready_sft.jsonl` | 201,756 | Yes |
| `validation_sft.jsonl` | 10,839 | No; checkpoint selection only |
| `research_test_eval.jsonl` | 11,171 | No; final test |
| `ready_cpt.jsonl` | 4,238 | Continued pretraining only |

`ready_sft.jsonl` contains 113,269 prior records plus
88,487 NuosuBench train records.

The three NuosuBench projections are physically separate and have disjoint
`source_row` sets. Because all three originate from the upstream public test
split, results must be reported as a same-source academic research split, not
as an uncontaminated official NuosuBench score.

All records retain upstream source, quality, and transformation metadata. The
large dictionary component remains marked as unreviewed community data.

## Citation

```bibtex
@dataset{axi_2026_nuosu_corpus,
  author    = {Wuhe Axi},
  title     = {Nuosu Task-Oriented Corpus},
  year      = {2026},
  version   = {2026.08.04},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/NiceAsiv/nuosu-corpus}
}
```