README.md

metadata

license: cc-by-4.0
pretty_name: RetroEnv-RL
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: dev
        path: tasks-public/dev.jsonl
      - split: test_id
        path: tasks-public/test_id.jsonl
      - split: test_hard
        path: tasks-public/test_hard.jsonl
  - config_name: tasks_with_known_routes
    data_files:
      - split: train
        path: tasks-private/train.jsonl
      - split: dev
        path: tasks-private/dev.jsonl
      - split: test_id
        path: tasks-private/test_id.jsonl
      - split: test_hard
        path: tasks-private/test_hard.jsonl

RetroEnv-RL

Multistep retrosynthesis tasks for agentic RL. Each task gives a target molecule, a depth budget (longest linear sequence) and optionally a constraint; the agent plans routes with tools and submits synthesis trees whose every leaf must be in the frozen stock. A deterministic verifier judges each step against a frozen reaction library (known reactions and frequent rdchiral retro-templates), never against a hidden answer; known routes only add a similarity bonus.

Split	Targets	Tasks	Use
train	70,028	75,224	RL rollouts and SFT bootstrap
dev	1,000	1,079	checkpoint selection, reward calibration
test_id	1,000	1,083	unseen targets, training distribution
test_hard	1,000	1,119	longer, novel, convergent or rare chemistry

Variants: max_depth (cap at the shortest known depth while the patent route is longer), restricted_stock (building blocks of both known routes removed), forbidden_class (a reaction class used by both known routes forbidden) and diversity (two or three routes with different first disconnections). Every task ships a witness route proving it is solvable under its constraint.

Held-out targets share no patent, route molecule, reaction, scaffold group or Tanimoto ≥ 0.90 near-duplicate with train or with each other; library reactions touching any held-out key are hidden from the agent's tools (library/reactions.jsonl.gz, visible).

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: every leaf of every PaRoutes route
library/	343,665 corpus reactions, 28,619 retro-templates, reagent whitelist
manifest.json, audit.json, checksums.json	build provenance, leakage and solvability audit, SHA-256
runs/	model evaluation episodes (full trajectories and scores)

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.