AdithyaSK/RetroEnv-RL

RetroEnv RL tasks

Tasks for RetroEnv, a multi-turn tool-use environment for retrosynthesis. Given a target molecule, the agent plans a synthesis back to purchasable building blocks: it searches the stock and training precedents, checks proposed disconnections, and submits route trees with emit_routes. A deterministic verifier scores the trees against the route reported in the target's patent, with a reward in [0, 1] built from nine components.

Code, the OpenEnv server, evaluation and training recipes: FineEnvs 09-retroenv (PR #30). SFT trajectories for the v3 train split: AdithyaSK/RetroEnv-SFT.

What is here

Path	Contents
tasks-public/	v3 tasks as the policy sees them: target, step budget, routes asked for, public difficulty
tasks-private/	the same tasks with their reference routes, the answer key the server grades against
stocks/paroutes-v2-n1.smi	the 13,432 purchasable molecules; stock_retrieve is the policy's only access to it
difficulty.jsonl	per-task heuristic tier and features (first reaction family, nearest train target, leaf rarity)
manifest.json, audit.json, checksums.json	how the splits were designed, the leakage audit, and SHA-256 of every file the server reads
normalized-routes.jsonl, baselines/report.json	every reference route, and oracle, one-route and empty-submission scores
retroeval-v2/	the earlier 1,000-task benchmark (2- and 3-step routes), same layout, with the model board in results/
runs/v2-eval/	the six models' v2 board episodes with full transcripts
media/retroenv-explainer.mp4	an 83-second explainer of one episode

v3 splits

Split	Tasks	2 / 3 / 4 / 5 steps	Two-route	Heavy atoms ≤20 / 21–30 / 31–40 / 41+	Easy / medium / hard
train	27,489	13,376 / 8,878 / 3,724 / 1,511	47	11,134 / 11,386 / 4,705 / 264	17,568 / 7,132 / 2,789
dev	150	42 / 42 / 36 / 30	15	42 / 51 / 45 / 12	60 / 56 / 34
eval	250	70 / 70 / 60 / 50	25	70 / 85 / 75 / 20	89 / 106 / 55
stress	150	42 / 42 / 36 / 30	15	35 / 51 / 47 / 17	58 / 56 / 36

Use train for RL, dev for checkpoint selection, eval for the board, and keep stress as a sealed second test. Eval covers all ten first-step reaction families (13 to 48 tasks each).

How the splits were made:

• The held-out splits were drawn first, against quotas on route length, first-step reaction family, target size and two-route tasks.
• No held-out task shares a target, scaffold, intermediate, reaction, patent or Morgan near-duplicate (Tanimoto ≥ 0.90) with train or with another held-out task.
• One-ring scaffolds, and the 33 multi-ring scaffolds found in at least 100 pool tasks (biphenyl, indole and others), group by exact structure. manifest.json lists them.
• A scripted solver that knows each route passes all 550 held-out tasks within the 16-turn, 32-call budget.

Use

Serve it with the OpenEnv server. This is the Docker image's default:

RETROENV_TASKS_REPO=AdithyaSK/RetroEnv-RL python envs/retro_route/openenv/prepare.py
# RETROENV_TASKS_SUBDIR=retroeval-v2 serves v2 instead

Load the rows:

from datasets import load_dataset

tasks = load_dataset("AdithyaSK/RetroEnv-RL", "v3")   # what the policy sees
answers = load_dataset("AdithyaSK/RetroEnv-RL", "v3-with-references")

Get a full local copy for the evaluation scripts and the explorer:

hf download AdithyaSK/RetroEnv-RL --repo-type dataset --local-dir benchmark/retroeval-v3 \
  --exclude "retroeval-v2/*" "runs/*" "media/*"

v2 model board

150 eval tasks, one attempt each, full toolset:

Model	Pass@1 (95% CI)	Exact route	Reward	Refused	Cost
Claude Opus 5.5	0.560 [0.48, 0.64]	0.593	0.750	16%	$13.94
Claude Sonnet 5.5	0.300 [0.23, 0.38]	0.340	0.626	0%	$9.76
GPT-5.6 Sol	0.280 [0.21, 0.36]	0.320	0.705	0%	$23.25
DeepSeek V4.1 Flash	0.107 [0.07, 0.17]	0.127	0.436	0%	$5.03
GPT-5.6 Luna	0.073 [0.04, 0.13]	0.100	0.565	0%	$2.29
Qwen3.8 27B	0.013 [0.00, 0.05]	0.007	0.138	0%	$8.27

A refusal ends the episode and scores the 0.05 floor. No models have been run on v3 yet.

Notes

• The answers to every split are public here, because PaRoutes is public and anyone can rebuild them. Say so when reporting results on a model that may have been trained on USPTO-derived data.
• validate_disconnection and reaction_class_lookup answer from the hidden reference. The server's unaided toolset removes them.

License and attribution

CC-BY-4.0. Routes and stock come from PaRoutes v2 (Genheden & Bjerrum, Digital Discovery 2022; Zenodo 7341155, CC-BY-4.0), which extracts reactions from USPTO patents.