Reaction-yield estimation

Answer one scientific question: for a fully specified reaction string, what yield does a trained regression model predict? Only after the exact deployment passes its canaries and held-out validation may the number rank comparable in-domain reactions or prioritize experiments. Do not call it a calibrated probability of step success, and never multiply step predictions into a route success probability.

Use sagawa/ReactionT5v2-yield, a 2025 MIT checkpoint trained on Open Reaction Database records and distributed with a direct local inference example. The model takes reactants, reagents, and product; a target alone is not valid input.

Deployment status: the currently pinned released checkpoint is quarantined for quantitative use. With the upstream wrapper, canonicalization, sorted mixture components, fixed 400-token padding, and the upstream Transformers version, the published model-card canary was expected to return about 19.1666% but returned 65.924858%. Until that discrepancy is resolved against a deployment-matched held-out set, the backend may be exercised for protocol testing only and its values must not rank reactions or support scientific conclusions.

Install and run

This Skill is self-contained; it does not require access to the reaction-forward-prediction Skill. Create the isolated environment with all direct and batch dependencies:

conda create -n reactiont5 python=3.11 -y
conda run -n reactiont5 python -m pip install \
  "torch" "transformers==4.40.2" "tokenizers==0.19.1" \
  "huggingface_hub[cli]==0.35.0" \
  sentencepiece rdkit datasets accelerate pandas

From an operator terminal whose current directory is the writable session workspace, acquire the reviewed source commit and immutable yield-model snapshot. Do not replace either revision with main; a different revision requires a new review and provenance record.

set -eu

REACTIONT5_ROOT="$PWD/models/reactiont5"
SOURCE_COMMIT="76eb08068e10fe255cae5d563a91e1c1e9abac54"
YIELD_REVISION="f0658bfd360bceaaf560f11b850781c50221fe0b"

mkdir -p "$REACTIONT5_ROOT"
if [ ! -d "$REACTIONT5_ROOT/source/.git" ]; then
  git clone https://github.com/sagawatatsuya/ReactionT5v2.git \
    "$REACTIONT5_ROOT/source"
fi
git -C "$REACTIONT5_ROOT/source" cat-file -e "${SOURCE_COMMIT}^{commit}"
git -C "$REACTIONT5_ROOT/source" checkout --detach "$SOURCE_COMMIT"
test "$(git -C "$REACTIONT5_ROOT/source" rev-parse HEAD)" = "$SOURCE_COMMIT"
SOURCE_STATUS="$(git -C "$REACTIONT5_ROOT/source" status \
  --porcelain --untracked-files=all)"
test -z "$SOURCE_STATUS"

conda run -n reactiont5 hf download sagawa/ReactionT5v2-yield \
  --revision "$YIELD_REVISION" \
  --local-dir "$REACTIONT5_ROOT/yield-$YIELD_REVISION"

The final assertion must remain empty; if a reused checkout has modified or untracked files, stop instead of importing it as reviewed source.

Record both revisions and hashes of the downloaded regular files, and keep the snapshot outside version control. Select the environment in its own OpenAI4S Python Cell:

host.env.use("reactiont5")

After the switch succeeds, import the reviewed wrapper from the pinned local source checkout and load only the reviewed local snapshot in a new Cell:

import os
import sys
from pathlib import Path

import torch
from transformers import AutoTokenizer

source = Path.cwd() / "models" / "reactiont5" / "source"
reviewed_revision = "f0658bfd360bceaaf560f11b850781c50221fe0b"
snapshot = Path.cwd() / "models" / "reactiont5" / f"yield-{reviewed_revision}"
if not source.is_dir() or not snapshot.is_dir():
    raise FileNotFoundError("reviewed ReactionT5 source or yield snapshot is missing")
os.environ["HF_HUB_OFFLINE"] = "1"
sys.path.insert(0, str(source))
from models import ReactionT5Yield2

model = ReactionT5Yield2.from_pretrained(snapshot, local_files_only=True)
tokenizer = AutoTokenizer.from_pretrained(snapshot, local_files_only=True)
model.eval()
text = "REACTANT:<reactants>REAGENT:<reagents>PRODUCT:<product>"
inputs = tokenizer([text], return_tensors="pt")
with torch.inference_mode():
    raw_predicted_percent = float(model(inputs).detach().cpu().reshape(-1)[0])
display_percent = min(100.0, max(0.0, raw_predicted_percent))

Do not load this regression checkpoint as a plain seq2seq model. Record the model ID, reviewed revision, local file hashes, source commit, package versions, device, and input string. Never fall back from a missing snapshot to a moving Hub model ID.

The upstream task_yield/prediction_with_PreTrainedModel.py script is not audit-compliant unchanged: it overwrites its prediction column with values clipped to 0–100. If adapting it for batches, preserve two columns before writing the CSV:

test_ds["prediction_raw"] = prediction
test_ds["prediction_percent"] = test_ds["prediction_raw"].clip(0, 100)

Run only that reviewed adaptation from the checkout's task_yield directory; never relabel the clipped column as the raw model result. Pass its --model_name_or_path argument the local models/reactiont5/yield-f0658bfd360bceaaf560f11b850781c50221fe0b directory and set HF_HUB_OFFLINE=1; do not pass the Hub model ID.

For OpenAI4S, use reaction_model_deployment.py to install the pinned shared ReactionT5v2 environment, download sagawa/ReactionT5v2-yield at revision f0658bfd360bceaaf560f11b850781c50221fe0b, snapshot the complete local model, and call ReactionModelBackend("reactiont5_yield", ...). The committed worker contains the model-card regression head, requires a local checkpoint, disables implicit downloads, preserves raw un-clipped output, and reports package and manifest provenance.

The worker also reproduces the pinned upstream preprocessing: each molecular mixture is RDKit-canonicalized component-wise, components are sorted, an absent reagent is encoded as one blank character, and inputs are padded/truncated to 400 tokens by default. input_max_length is recorded and bounded to 32--1024. Matching preprocessing did not remove the canary discrepancy above.

Copy the wrapper exactly from the official model card or repository rather than loading the checkpoint as a plain seq2seq model. Pin the Hugging Face revision and record package versions, device, input string, and checkpoint hash.

Scenario 6 benchmark contract

Use ../retrosynthesis_planning/yield_benchmark.py for the frozen random test and four molecular-framework OOD groups. Submit the raw predicted percentage; never clip it before evaluation. Intervals must be either fully specified or explicitly absent, and every prediction carries a domain-status label. The evaluator reports per-group MAE/RMSE/R2/ranking and interval diagnostics, macro-OOD MAE, and worst-group MAE rather than hiding shift behind one pooled score.

Domain gate

Before quoting the number, record:

• whether reagents, catalyst, solvent, and temperature are known or missing;
• whether the reaction class and substrate family resemble the validation data;
• whether the value comes from the base checkpoint or a deployment-specific fine-tune;
• held-out MAE/RMSE and calibration diagnostics for that deployment domain;
• an uncertainty estimate, if and only if one was actually computed by a validated ensemble or conformal procedure.

If these checks are absent, label the output screening_only. The published benchmark includes strong C-N coupling results, but that does not establish uniform accuracy across arbitrary chemistry or laboratory protocols.

Output contract

Return reaction fields, predicted yield percent, raw unclipped value, model and revision, domain status (matched, uncertain, out_of_domain), missing-input flags, optional validated uncertainty interval, and evaluation provenance. Clip only for presentation; preserve any raw prediction outside 0–100 for audit.

Failure modes

Symptom	Action
product or reagent context missing	Refuse quantitative interpretation; request a complete reaction record.
raw prediction outside 0–100	Preserve it, flag extrapolation, and show a clipped display value only if needed.
released model-card canary mismatch	Quarantine the checkpoint; do not rank reactions until independently resolved.
no deployment-matched held-out set	Label screening_only; do not state expected experimental error.
multiple route steps	Score steps separately and report the weakest/most uncertain steps; never multiply percentages.

Primary sources: https://github.com/sagawatatsuya/ReactionT5v2 and https://huggingface.co/sagawa/ReactionT5v2-yield.