Datasets:

UniParser
/

RxnMiner-Dataset

pretty_name: RxnMiner Dataset (reactions extracted from RxnCaption-15k)
license: cc-by-nc-sa-4.0
language:
  - en
tags:
  - chemistry
  - reaction-extraction
  - reactions
  - silver-standard
  - agent-trajectories
task_categories:
  - image-to-text
size_categories:
  - 10K<n<100K
configs:
  - config_name: default
    default: true
    data_files:
      - split: train
        path: data/train-*.parquet
  - config_name: trajectories
    data_files:
      - split: train
        path: trajectories/train-*.parquet

RxnMiner Dataset

129,778 concrete reactions and 15,950 generic templates extracted by RxnMiner, without human intervention, from 15,000 literature reaction figures (the training split of RxnCaption-15k). Every record is linked to the figure it came from, and the full agent session that produced it is released alongside.

This is a silver dataset: it is machine-extracted and has not been reviewed. For reference, RxnMiner with the same backbone scores 71.0 macro reaction F1 (stereochemistry-aware) on the expert-reviewed benchmark UniParser/RxnMiner-Benchmark; that score is measured on a different set of figures and is not an estimate of the error rate of this dataset. No image in this dataset is byte-identical to a benchmark figure (compared by SHA-256).

It accompanies the paper RxnMiner: A Multimodal Agent for Structured Reaction Data Extraction from Chemistry Figures (arXiv:2611.02826). Code: github.com/rxnminer/RxnMiner.

The paper describes an earlier run over the RxnCaption-11k training split (10,107 figures, 88,202 reaction records). This release is a later run of RxnMiner over all 15,000 figures of the RxnCaption-15k training split, so its counts differ from those reported in the paper.


Figures	15,000 (all runs finished with a complete answer)
Concrete reactions	129,778
Generic templates	15,950
Model	Qwen3.8-Flash-Next, deployed locally; one agent run per figure
Run cost	717,187 model calls; 53.6 B input tokens, 0.85 B output tokens; median 17.5 min per figure (from `elapsed_sec`)

Configurations

default (5.2 GB): one row per figure.

Field	Type	Description
`id`	string	Figure id (the RxnCaption-15k image name without extension)
`image`	image	The source figure
`extraction`	string (JSON)	RxnMiner's answer in the unified JSON format: `selected_examples` = concrete reactions, `reaction_templates` = generic schemes with R groups, `reaction_graph` = multistep links. Each molecule carries `smiles` and the byte-exact `E-SMILES`; `_source` records how the structure was obtained
`plan`	string (JSON)	The plan the agent wrote for the figure before extracting
`num_reactions`	int	Concrete reactions in `selected_examples`
`num_templates`	int	Generic templates in `reaction_templates`
`is_multistep`	bool	The figure shows a multistep route
`answer_complete`	bool	The run ended with a complete answer
`model`	string	Model behind the agent
`elapsed_sec`	float	Wall-clock time of the run
`tool_calls`, `llm_calls`	int	Tool calls and model calls in the run
`input_tokens`, `output_tokens`, `reasoning_tokens`	int	Token usage of the run
`source_dataset`	string	Always `RxnCaption-15k (train_set)`

trajectories (27 GB): id and trajectory, the full agent session as JSON lines (every model message, tool call and tool result). Credentials and internal hostnames, IP addresses and file paths that appeared in tool output were replaced with [REDACTED], [INTERNAL-IP] or [INTERNAL-PATH] (103 of the 15,000 sessions were affected). Unlike the benchmark evaluation reported in the paper, which ran without external network access, these corpus runs had internet access. In fewer than 100 of the 15,000 sessions the agent queried public chemistry databases (for example PubChem) or literature search services; those calls are visible in the trajectories.

import json
from datasets import load_dataset

ds = load_dataset("UniParser/RxnMiner-Dataset", split="train")
record = json.loads(ds[0]["extraction"])
for rxn in record["selected_examples"]:
    print([m.get("smiles") for m in rxn["reactants"]], ">>", [m.get("smiles") for m in rxn["products"]])

# Agent sessions are large; stream them instead of downloading all 27 GB.
traj = load_dataset("UniParser/RxnMiner-Dataset", "trajectories", split="train", streaming=True)

License and attribution

The extracted records and trajectories are released under CC BY-NC-SA 4.0. The figures come from the training split of RxnCaption-15k (arXiv:2511.02384) and remain subject to its terms and to the rights of the original publishers. Please cite RxnCaption when you use the figures, and the RxnMiner paper when you use the extracted records or trajectories.

Citation

@misc{hu2026rxnminer,
  title         = {RxnMiner: A Multimodal Agent for Structured Reaction Data Extraction from Chemistry Figures},
  author        = {Hu, Jinbo and Zhang, Songren and Fu, Chuanxun and Huang, Jinrui and Wong, Zhen Hao and Cui, Dongxu and
                   Fan, Haoran and Ge, Muhao and Jian, Zhen and Zhou, Jian and Luo, Chengxiang and Lyu, Han and Tao, Haoyi and
                   Huang, Chaozheng and Li, Hanzheng and Wang, ZiChen and Ma, Hengtian and Dai, Jinyang and Wang, Nan and
                   Hong, Yanhui and Fang, Xi and Ke, Guolin},
  year          = {2026},
  eprint        = {2611.02826},
  archivePrefix = {arXiv},
  url           = {https://arxiv.org/abs/2611.02826}
}