README.md

metadata

license: cc-by-4.0
pretty_name: RetroEnv-Curated
task_categories:
  - reinforcement-learning
  - text-generation
tags:
  - chemistry
  - retrosynthesis
  - tool-use
  - rl-environment
  - openenv
size_categories:
  - 10K<n<100K
configs:
  - config_name: tasks
    default: true
    data_files:
      - split: train
        path: tasks-public/train.jsonl
      - split: validation
        path: tasks-public/validation.jsonl
      - split: heldout
        path: tasks-public/heldout.jsonl
      - split: test
        path: tasks-public/test.jsonl
  - config_name: tasks_with_known_routes
    data_files:
      - split: train
        path: tasks-private/train.jsonl
      - split: validation
        path: tasks-private/validation.jsonl
      - split: heldout
        path: tasks-private/heldout.jsonl
      - split: test
        path: tasks-private/test.jsonl

RetroEnv-Curated

Four splits for training and evaluating retrosynthesis agents, cut from LiteFold/RetroEnv. Each task gives a target molecule, a depth budget and optionally a constraint; the agent plans with tools and submits synthesis trees whose every leaf must be in the frozen stock. A deterministic verifier judges each step against the reaction library, never against a hidden answer.

This is the whole dataset. There are no alternative subsets to pick between.

Split	Tasks	easy / medium / hard	Use
train	13,200	6812 / 5153 / 1235	RL rollouts and SFT
validation	150	50 / 50 / 50	checkpoint selection during training
heldout	300	100 / 100 / 100	sealed final test; run once
test	150	50 / 50 / 50	the iteration board

The four splits never share a task. Held-out targets share no patent, route molecule, reaction, scaffold group or Tanimoto >= 0.90 near-duplicate with train or with each other, and library reactions touching a held-out key are hidden from the agent's tools. Within a tier the draw is spread over shortest-route depth and prefers a first reaction the tier has not used yet, so even the 150-task splits cover the chemistry rather than one corner of it.

train carries 1,200 constraint variants on top of its standard tasks (max_depth, restricted_stock, forbidden_class, diversity), because those rules are only learnable from tasks that impose them. train is also the one split name that must not change: the environment hides a task's own reactions and precedents from the tools only for tasks whose split is train.

Path	Contents
tasks-public/	what a policy may see
tasks-private/	the same tasks with hidden known routes and difficulty labels
stocks/paroutes-archive-leaves.smi	102,915 building blocks
library/	corpus reactions, retro-templates, reagent whitelist
manifest.json, checksums.json	build provenance and SHA-256

Use

hf download LiteFold/RetroEnv-Curated --repo-type dataset --local-dir RetroEnv-Curated
RETROENV_BENCHMARK_DIR=RetroEnv-Curated uv run uvicorn retroenv_openenv.server:app --port 8000

Source: PaRoutes v2 (Genheden & Bjerrum, Zenodo 7341155, CC-BY-4.0), patent reactions from USPTO. Known routes are public, so report results knowing a model may have seen them.