Datasets:

Mingzhou9
/

sarl

You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

SA-RL USPTO

Private USPTO chemical structure recognition dataset for the SA-RL project.

Repository layout

uspto/images/: 30,000 molecular structure images. Files already committed before Hub sharding remain directly under images/; the rest are distributed under two-character hash subdirectories. Annotation images fields contain the exact relative path.
uspto/sft_difficulty.json: 30K three-stage SFT annotations with difficulty metadata.
uspto/gt_cot_sft_difficulty.json: 30K rule-grounded observation annotations with difficulty metadata.
uspto/sft_difficulty_5k.json: 5K difficulty-matched SFT subset.
uspto/llm_sft_difficulty_5k.json: 5K mixed LLM/rule observation subset.
uspto/dataset_info.json: LlamaFactory-style dataset registration metadata.
uspto/difficulty_config.json: difficulty scoring configuration.

The <observe> section describes visible atoms or vertices and their 2D and normalized coordinates before formal atom IDs are assigned in Stage 1.