---
license: other
license_name: mixed
license_link: LICENSE
language:
- en
- zh
- multilingual
task_categories:
- text-generation
- image-text-to-text
tags:
- reasoning
- chain-of-thought
- multimodal
- grounding
size_categories:
- 1M<n<10M
---

# thoughtbase-v2

Text and image prompts, each with one or more reasoning traces and a final response, collected from
existing open datasets. Nothing here was generated for this release; rows were filtered, deduplicated
and regrouped. Built for non-commercial research.

| | count |
|---|---|
| prompts (train / validation) | 2,740,885 / 27,510 |
| reasoning traces (all responses) | 4,722,758 |
| image traces | 934,278 (19.8%) |
| image files | 45 tar shards under `images/` |

## Format

`train.jsonl` and `validation.jsonl`, one prompt per line:

```json
{"id": "v3_gqa_1:…", "prompt": "What is the color of that vase?…",
 "image_paths": ["images/ab/ab12….jpg"],
 "responses": [{"cot": "The user wants…", "final_answer": "Green", "model": "v3_gqa_1"}],
 "split_group": "image:ab12…", "source": {"dataset": "nvidia/Nemotron-Image-Training-v3", "source": "v3_gqa_1", "row": 17}}
```

- `responses` holds every distinct trace for the same prompt and images (e.g. several solutions of one
  math problem). `model` names the source the trace came from.
- Multi-turn source rows become one prompt per turn: earlier turns keep their question and final
  answer, only the target turn has reasoning.
- Grounding answers use `<box>[[x1, y1, x2, y2], …]</box>` with coordinates on a 0–1000 grid.
- About 1% of prompts are in validation, split by image hash or whitespace-normalized prompt, so no
  image or prompt appears in both splits.
- Length limits (Qwen3.5 tokenizer, real token counts): prompt + 256 tokens per image ≤ 4096,
  cot < 16384, final answer < 4096. Longer items were dropped, not truncated.

## Use

```bash
hf download JonesLin/thoughtbase-v2 --repo-type dataset --local-dir thoughtbase-v2
cd thoughtbase-v2 && for t in images/*.tar; do tar -xf "$t"; done   # creates images/<sha[:2]>/<sha>.<ext>
```

Image paths in the JSONL are relative to the folder holding it.

## Sources

`manifest.json` has prompts, responses and drops per source. Main sources (prompts / traces):

| group | source | license of the source |
|---|---|---|
| math | nvidia/Nemotron-Math-v2 (low, medium) | CC-BY-4.0 / CC-BY-SA-4.0 |
| math, code, Q&A | PrimeIntellect/SYNTHETIC-1 (verified score) | Apache-2.0 |
| code, IF, chat, safety, multilingual | allenai/Dolci-Think-SFT-7B | ODC-BY-1.0 |
| code | open-r1/codeforces-cots | CC-BY-4.0 |
| science | nvidia/Nemotron-Science-v1, nvidia/OpenScienceReasoning-2 | CC-BY-4.0 |
| medical | FreedomIntelligence/medical-o1-reasoning-SFT | Apache-2.0 |
| instruction following | nvidia/Nemotron-SFT-Instruction-Following-Chat-v3 | mixed, see source |
| multilingual | nvidia/Nemotron-SFT-Multilingual-v2 | CC-BY-4.0 / CC-BY-SA-4.0 |
| chat | nvidia/Llama-Nemotron-Post-Training-Dataset | CC-BY-4.0 |
| images | nvidia/Nemotron-Image-Training-v3 (aokvqa_2, gqa, chartqa, plotqa, figureqa, clevr, mulberry_1/2, flickr30k, toloka, zhwiki_1, unigeo, geometry3k, pmc_vqa, slake_med) | CC-BY-4.0 annotations; ChartQA labels GPL-3.0; zhwiki CC-BY-SA-4.0 |
| images | Fancy-MLLM/R1-Onevision | Apache-2.0 |
| images | remyxai/SpaceThinker (spatial reasoning) | Apache-2.0 |
| images | HuggingFaceM4/A-OKVQA, derek-thomas/ScienceQA, LightChen2333/M3CoT | not stated / CC-BY-SA-4.0 / MIT |

Images come from the original releases (COCO, GQA, Flickr30k, ChartQA, FigureQA, CLEVR, PMC-VQA,
SLAKE, UniGeo, Geometry3K, Mulberry-SFT and others) and keep their own terms. **Flickr30k images are
for non-commercial research only.** Check each source's license before any other use.

mulberry_2 traces were kept only when their final answer matched Mulberry-SFT's answer.