README.md

---
pretty_name: TRACE
language:
- en
license: cc-by-4.0
annotations_creators:
- machine-generated
source_datasets:
- original
task_categories:
- visual-question-answering
size_categories:
- 10K<n<100K
tags:
- visual-reasoning
- synthetic
- rlvr
- multimodal
- grounded-reasoning
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train/trace_rlvr_train_64000_all1000_seed42-*.parquet
  - split: validation
    path: data/validation/trace_rlvr_validation_iid_2000_all1000_seed1042.parquet
---

# TRACE: A Taxonomy-Guided Environment for Multidomain Visual Reasoning

TRACE is a synthetic dataset for grounded visual reasoning and reinforcement
learning with verifiable rewards. Each example pairs a rendered image with two
prompt variants, typed targets, a reward contract, and a reference to its
execution trace.

[Paper](https://arxiv.org/abs/2607.19790) ·
[Project page](https://maveryn.github.io/trace/) ·
[GitHub](https://github.com/maveryn/trace) ·
[Collection](https://huggingface.co/collections/maveryn/trace-6a604291b4be4ed6399b9f24) ·
[Task catalog](https://github.com/maveryn/trace/blob/main/docs/TASK_CATALOG.md) ·
[Training recipes](https://github.com/maveryn/trace/tree/rlvr/rlvr) ·
[3B model](https://huggingface.co/maveryn/trace-qwen2.5-vl-3b) ·
[7B model](https://huggingface.co/maveryn/trace-qwen2.5-vl-7b)

## Dataset structure

The release covers 1,000 tasks, 277 scenes, and 11 visual domains: charts,
games, geometry, graphs, icons, illustrations, pages, physics, puzzles,
symbolic notation, and 3D scenes.

| Split | Examples | Coverage | Generation seed |
| --- | ---: | --- | ---: |
| train | 64,000 | 64 examples per task | 42 |
| validation | 2,000 | 2 examples per task | 1042 |

Both splits are deterministically shuffled with row-order seed `20260711`.
Validation is IID over the same 1,000 tasks; it is not a held-out test of new
task families.

This is the canonical dataset release for reproducing the public TRACE 3B and
7B checkpoints. The `dataset-v1` tag is the stable human-facing release name;
the public training configurations pin its immutable commit. An independent
[equivalence receipt](metadata/reproducibility/equivalence_receipt.v1.json)
verifies that every field consumed by training is identical to the original
run input, including embedded image bytes and row order.

The optional `trace_supervision_mode` column is advisory metadata. The released
training profiles select `prompt_answer`, score `answer_gt`, and do not read
this column.

## Fields

| Field | Description |
| --- | --- |
| `images`, `image_sizes` | Embedded rendered image and dimensions defining the annotation coordinate frame |
| `prompt_answer` | Prompt selected by the released training profiles |
| `prompt_answer_and_annotation` | Prompt requesting an answer and image-space annotation |
| `answer_gt`, `annotation_gt` | JSON-encoded typed verifier targets |
| `reward_contract` | JSON-encoded metadata-backed scoring contract |
| `instance_id`, `domain`, `scene_id`, `task`, `query_id` | Identity and public taxonomy |
| `trace_ref` | JSON-encoded reference to the compressed execution-trace sidecar |
| `trace_supervision_mode` | Optional advisory prompt-selection recommendation; unused by the released training profiles |

The `sidecars/` directories contain build and validation reports, canonical
instance records, curriculum indices, manifests, and compressed execution
traces. Images are embedded in the Parquet files and are not duplicated in
the sidecars. Report and manifest paths on the default branch are
repository-relative; dataset ids and content hashes retain artifact identity.

## Loading

```python
from datasets import load_dataset

dataset = load_dataset("maveryn/trace", revision="dataset-v1")
train = dataset["train"]
validation = dataset["validation"]
```

The contract fields are stored as JSON strings. Decode them before scoring:

```python
import json

row = dataset["train"][0]
answer = json.loads(row["answer_gt"])
annotation = json.loads(row["annotation_gt"])
reward_contract = json.loads(row["reward_contract"])
```

The public [3B](https://github.com/maveryn/trace/blob/rlvr/rlvr/configs/trace-qwen2.5-vl-3b.yaml)
and [7B](https://github.com/maveryn/trace/blob/rlvr/rlvr/configs/trace-qwen2.5-vl-7b.yaml)
training configurations pin the immutable release commit and validate the split
sizes, task coverage, schema, prompts, and model inputs before training.

## Intended use and limitations

TRACE is intended for research on multimodal reasoning, verifiable-reward
post-training, grounded supervision, synthetic curricula, and task-level
analysis. Its procedurally generated images do not represent the full
distribution of natural images or real documents. Prompts are predominantly
English, and metadata-backed labels do not eliminate model or task-design
biases. The dataset should not be treated as evidence of real-world model
reliability or safety.

## Citation

If you use TRACE, please cite:

```bibtex
@misc{alam2026trace,
  title         = {Trace: A Taxonomy-Guided Environment for Multidomain Visual Reasoning},
  author        = {Alam, Md Tanvirul},
  year          = {2026},
  eprint        = {2607.19790},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2607.19790}
}
```

## License

The dataset is released under
[CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). The generator
software is separately licensed under
[Apache-2.0](https://github.com/maveryn/trace/blob/main/LICENSE); bundled assets
and dependencies retain the terms documented in the source repository.