AdithyaSK/RetroEnv-SFT

RetroEnv SFT trajectories

Multi-turn tool-calling episodes for RetroEnv, a retrosynthesis environment. Each row plans a route for one v3 train task of AdithyaSK/RetroEnv-RL back to purchasable molecules and ends with emit_routes. Every row is an episode that the environment's verifier passed, through the same OpenEnv server and agent loop used for evaluation.

	chemist (default)	plain (baseline)
Rows (train / validation)	31,897 / 323	27,222 / 267
Train tasks covered	27,489	27,489
Rows with 4- or 5-step routes	31%	19%
Exact match to the patent route	96.8%	96.8%
Rows that recover from a rejected cut	46%	0%
Tools used	5	3
Mean tool calls per row	11.8	8.6
Distinct assistant texts	112,902	31,736
Tokens per row, p50 / p95 (Qwen3.5)	5,473 / 7,778	4,142 / 5,858

Use chemist. plain is the first version, kept for comparison: an expert that is always right, writes four fixed sentences, and never recovers from a mistake.

How chemist rows were made

A scripted chemist that knows each task's patent route plays the episode through the real server:

1. Most episodes start by inspecting the target and searching the training precedents for close analogues.
2. It proposes disconnections from a library of strategic bonds: amide, sulfonamide, ester, ether, biaryl, and C–N by alkylation, reductive amination or aryl substitution. In 46% of episodes at least one proposal is a plausible cut the patent did not make. validate_disconnection rejects it, and the chemist moves on.
3. Each cut is validated and every new piece is looked up in the stock. A piece the stock lacks gets its own disconnection; one it holds stays a leaf, even where the patent made it, so some routes are shorter than the reference.
4. The submitted tree names each reaction, its reagent roles and its evidence.

Its text states only what tool results returned or what the structures show: reaction families, reagent roles, strategic bonds. It cites a precedent only when its reaction type matches, and never writes the task's own patent number, reaction IDs or reagents. Choices are seeded per episode. 4- and 5-step tasks keep up to two different trajectories each, to move the mix toward the eval split. 1% of tasks are held out as validation.

No row comes from a dev, eval or stress task. The export refuses them, and re-audits the train tasks used against the held-out splits under the benchmark's own leakage rules.

Row layout

messages is an OpenAI-style chat: a system prompt, the task prompt, then assistant tool calls and tool results, ending with the assistant's emit_routes call. Tool results are verbatim server output, including model_turns_remaining. tools holds the nine function schemas. Other columns: id, task_id, target_smiles, max_steps, kind, reward, exact_match, tool_calls, turns, toolset, source, provider.

Tool-call arguments and tools are JSON strings, so that Arrow does not merge different tools' arguments into one struct with null-filled gaps. Decode them before apply_chat_template:

import json
from datasets import load_dataset

rows = load_dataset("AdithyaSK/RetroEnv-SFT", "chemist", split="train")

def decode(row):
    messages = []
    for m in row["messages"]:
        m = {k: v for k, v in m.items() if v is not None}
        for call in m.get("tool_calls") or []:
            call["function"]["arguments"] = json.loads(call["function"]["arguments"])
        messages.append(m)
    return messages, json.loads(row["tools"])

train/sft_smoke.py in the RetroEnv code trains on this with TRL, with loss on assistant turns only.

Limitations

• The rejected cuts come from eight bond rules, so the mistakes are textbook alternatives rather than the varied errors a model makes.
• The text comes from short templates with variants: factual, not free-form reasoning.
• validate_disconnection answers from the task's hidden reference. These rows use the full toolset, so they teach a policy to lean on it.

License and attribution

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