EER6/nvidia-OpenCodeInstruct-broad

---
license: apache-2.0
source_datasets:
- nvidia/OpenCodeInstruct
task_categories:
- text-generation
language:
- en
tags:
- code
- sft
- instruction-tuning
- filtered
size_categories:
- 1M<n<10M
---

# nvidia-OpenCodeInstruct-broad

A quality-filtered subset of [nvidia/OpenCodeInstruct](https://huggingface.co/datasets/nvidia/OpenCodeInstruct) (5M examples).

## Filtering criteria

Both conditions must be satisfied:

| Criterion | Threshold |
|-----------|-----------|
| **LLM judge min score** | >= 4 (out of 5) |
| **Unit test pass rate** (`average_test_score`) | >= 0.8 |

**LLM judge min score** is the minimum across all three dimensions in the `llm_judgement` field:
- `requirement_conformance` — does the code do what the instruction asked?
- `logical_correctness` — is the algorithm/logic correct?
- `edge_case_consideration` — does it handle edge cases?

A min score >= 4 means *every* dimension scores at least 4/5.

**Unit test pass rate** (`average_test_score`) is the fraction of 10 LLM-generated unit tests that the solution passes. >= 0.8 means at least 8 out of 10 tests pass.

## Result

- **Source size:** 5,000,000
- **Filtered size:** 1,698,239
- **Retention rate:** 34.0%

[EER6/nvidia-OpenCodeInstruct-refined](https://huggingface.co/datasets/EER6/nvidia-OpenCodeInstruct-refined) is a strict subset of this dataset with tighter thresholds (llm_min = 5, test = 1.0, 444K examples).

## Schema

All original columns from `nvidia/OpenCodeInstruct` are preserved as-is — no transforms or column additions. See the [original dataset card](https://huggingface.co/datasets/nvidia/OpenCodeInstruct) for column descriptions and citation.