README.md

---
pretty_name: JupiMind datasets
language:
- vi
- en
license: other
size_categories:
- 1M<n<10M
task_categories:
- text-generation
tags:
- pretraining
- vietnamese
- technical-documentation
- mixed-license
source_datasets:
- original
---

# JupiMind datasets

This repository stores normalized dataset snapshots used for JupiMind training workflows. The primary asset at the moment is the pretraining mix built for `aevynt/JupiMind-Dense-J1R-Tokenizer-24k`.

## Files currently included

- `pretrain/j1r_24k_7b_mix/j1r_t4x4_24k_mix.jsonl`
- `pretrain/j1r_24k_7b_mix/j1r_t4x4_24k_mix.summary.json`

## Main dataset: J1R 24k 7B mix

- Type: pretraining JSONL
- Format: one record per line as `{"text": "..."}`
- Target tokenizer: `aevynt/JupiMind-Dense-J1R-Tokenizer-24k`
- Generated sequences: `1,710,000`
- Target context length: `4096`
- Estimated token budget: about `~7B` tokens using chunk size `4094`
- Original build output path: `/home/trainer/jupimind/outputs/pretrain/j1r_t4x4_24k_mix.jsonl`

## Actual source mix after backfill

Two smaller technical sources did not contain enough material to reach their planned quotas, so the build pipeline backfilled the missing volume with `fineweb2_vi`. The table below reports the **actual final composition**, not the initial intended weights.

| Source key | Upstream dataset | Config | Written sequences | Actual share | Initial target | Source shortfall | Backfill |
| --- | --- | --- | ---: | ---: | ---: | ---: | ---: |
| `fineweb2_vi` | `epfml/FineWeb2-HQ` | `vie_Latn` | 781,741 | 45.72% | 564,300 | 0 | 217,441 |
| `vietnamese_curated` | `VTSNLP/vietnamese_curated_dataset` | `-` | 376,200 | 22.00% | 376,200 | 0 | 0 |
| `wikipedia_vi` | `wikimedia/wikipedia` | `20231101.vi` | 171,000 | 10.00% | 171,000 | 0 | 0 |
| `stackv2_edu` | `common-pile/stackv2_edu_filtered` | `-` | 256,500 | 15.00% | 256,500 | 0 | 0 |
| `github_ai_docs` | `meowterspace42/github-ai-project-docs` | `-` | 3,047 | 0.18% | 136,800 | 133,753 | 0 |
| `tech_docs` | `saidsef/tech-docs` | `-` | 1,812 | 0.11% | 85,500 | 83,688 | 0 |
| `fineweb_edu` | `HuggingFaceFW/fineweb-edu` | `sample-10BT` | 119,700 | 7.00% | 119,700 | 0 | 0 |

## Provenance and licensing

This repository does **not** apply a single blanket content license to the mixed dataset. Each source portion remains governed by its upstream terms. See [`LICENSE`](./LICENSE) and [`SOURCES.md`](./SOURCES.md) for details.

| Upstream dataset | Upstream license or terms | Link |
| --- | --- | --- |
| `epfml/FineWeb2-HQ` | ODC-By 1.0; also subject to CommonCrawl Terms of Use | [upstream](https://huggingface.co/datasets/epfml/FineWeb2-HQ) |
| `VTSNLP/vietnamese_curated_dataset` | No explicit Hub license field; review the upstream dataset card and underlying component datasets. | [upstream](https://huggingface.co/datasets/VTSNLP/vietnamese_curated_dataset) |
| `wikimedia/wikipedia` | CC-BY-SA 3.0 and GFDL | [upstream](https://huggingface.co/datasets/wikimedia/wikipedia) |
| `common-pile/stackv2_edu_filtered` | Open-source licenses vary by document; review upstream card and per-row metadata. | [upstream](https://huggingface.co/datasets/common-pile/stackv2_edu_filtered) |
| `meowterspace42/github-ai-project-docs` | Apache-2.0 per Hub metadata; original source repositories may still carry their own licenses. | [upstream](https://huggingface.co/datasets/meowterspace42/github-ai-project-docs) |
| `saidsef/tech-docs` | Apache-2.0 | [upstream](https://huggingface.co/datasets/saidsef/tech-docs) |
| `HuggingFaceFW/fineweb-edu` | ODC-By 1.0; also subject to CommonCrawl Terms of Use | [upstream](https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu) |

## Usage notes

- This is a **derived training mix** assembled for pretraining, not a replacement distribution for the original source datasets.
- Downstream users are responsible for reviewing legal, safety, privacy, attribution, and redistribution obligations of each upstream source before training or deploying a model.
- `VTSNLP/vietnamese_curated_dataset` and `common-pile/stackv2_edu_filtered` require extra care because their final content rights depend on upstream components or per-document metadata.

## Quick load

```python
from datasets import load_dataset

ds = load_dataset(
    "aevynt/jupimind-datasets",
    data_files={"train": "pretrain/j1r_24k_7b_mix/j1r_t4x4_24k_mix.jsonl"},
    split="train",
)
print(ds[0])
```