You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

ChemCompass v3 β€” normalized across all published source collections

V3 contains 43,573,294 source observations in 39 collections, including the full FREA, FREA legacy, CREED and CREED-CCV corpora. Every source observation passes through the same typed normalization pipeline. These totals include memberships, explicit derived variants, route documents and metadata; they are not counts of independent judgments or unique reactions.

Use v3_flat_<collection> for one observation per row, v3_grouped_<collection> for chemical entities with complete observation lists, or v3_flat_all to stream every collection. v3_grouped_all concatenates collection-scoped groups; the same entity may appear in several collections. Grouped rows preserve every score, explanation, provenance field and raw source object. Groups above 10,000 observations use group_part / is_last_part without discarding observations.

from datasets import load_dataset

frea = load_dataset("shreyasv/chemcompass", "v3_flat_frea_corpus", split="train", streaming=True)
creed = load_dataset("shreyasv/chemcompass", "v3_flat_creed", split="observations", streaming=True)
expert = load_dataset("shreyasv/chemcompass", "v3_grouped_frea_expert", split="entities", streaming=True)

FREA's flat configuration retains its original train/test splits. Grouping combines each collection's splits, with source_split retained on each observation. Do not use grouped FREA to construct a train/test split.

Chemical normalization uses map-free canonical isomeric RDKit SMILES and sorted disconnected components, retaining stereochemistry, isotopes, charges and fragments. No salt stripping or tautomer merging is performed. Invalid/unsupported chemistry and unresolved linkage remain explicit records with null normalized chemistry and preserved originals. Normalization status counts: {"invalid": 79, "not_applicable": 37515, "valid": 43535700}.

Native score scales are retained. CREED scores remain null; FREA corpus labels describe construction; CCV scores come from ChemCensor 0.5.2. Human, model, derived and unscored records are distinguishable. Route drawings lacking explicit reaction links retain normalized molecules and hierarchical annotations, without invented reactions. See the schema, exact recovered rubrics and version caveats, coverage inventory, validation receipt and its sampling boundaries, and source-specific licenses. CREED/CCV remain noncommercial.

Collection Observations
creed 22,712,526
creed_ccv 6,368,986
development_features 112
development_records 121
development_unlinked_steps 199
example_records 10
frea_corpus 14,115,057
frea_expert 751
frea_legacy 250,000
route_catalog_molecules 1,135
route_catalog_reactions 555
route_catalog_routes 171
route_expert_consensus 120
route_expert_individual 991
route_expert_whole_route 200
route_llm_claude_opus48 3,230
route_llm_claude_opus48_evaluations 800
route_llm_claude_opus48_whole_route 800
route_llm_gemini_35_pro 3,264
route_llm_gemini_35_pro_evaluations 800
route_llm_gemini_35_pro_whole_route 800
route_llm_gpt55 3,262
route_llm_gpt55_evaluations 800
route_llm_gpt55_whole_route 800
route_llm_llama31_70b 3,052
route_llm_llama31_70b_evaluations 800
route_llm_llama31_70b_whole_route 737
three_class_route_expert_consensus 116
three_class_route_expert_individual 838
three_class_route_llm_claude_opus48 2,781
three_class_route_llm_gemini_35_pro 2,636
three_class_route_llm_gpt55 2,581
three_class_route_llm_llama31_70b 2,892
ursa_chemcensor 1,000
ursa_chemcensor_binary 1,000
ursa_expert 1,000
ursa_routes 82,129
ursa_targets 6,192
validation_records 50

Prior releases and the original default configuration remain available below.


ChemCompass

Retrosynthesis route structures, expert assessments and model feedback in self-contained structured JSON files and flat exports. The structured files embed route context; the flat files contain reaction SMILES and complete assessments, with scoring definitions provided in this repository.

Expanded collections β€” October 2026

Schema v2 adds 751 FREA expert reactions and 1,000 URSA expert reactions, giving 2,862 expert assessments across 2,073 reactions. All original human and LLM components also have v2 flat and grouped views with native score scales, rubric IDs, assessment scope and source provenance. Existing JSON exports, three-class variants and the default configuration are retained. All HF configurations use Parquet; small-table Parquet files in tables/ are lossless equivalents of the retained flat/grouped JSON and older FREA CSV.

Separate corpora include the complete 14,115,057-row FREA corpus, its older 250,000-row release, CREED's 22,712,526 generated candidates and the 6,368,986-candidate CREED-CCV subset. Six URSA target files and original route archives are also included. Corpus construction labels and machine scores are not expert annotations; overlapping collections must not be summed as unique reactions. CREED/CREED-CCV carry a noncommercial license. See all source-specific licenses and pinned provenance.

Use v2_all_expert for grouped expert assessments, v2_flat_all_expert for flat assessments, and frea_corpus for streaming train/test Parquet. The sections below describe the original release and its original counts.

Additional three-class version

The three_class/ version applies to both human cohorts and all four LLMs: original scores 3/4 are removed, and 1 β†’ 1, 2 β†’ 2, 5 β†’ 3. It includes flat and grouped JSON, remapped rubrics with verbatim criteria, and coverage counts. Original scores and rubric IDs remain as provenance fields. Load it with configuration three_class_expert (or another three_class_ subset). The original five-point data and default configuration remain unchanged. The counts below describe the original data; variant counts are in its README.

Dataset counts

Measure Count
Distinct route structures 171
Reaction occurrences 555
Distinct canonical reactions 552
Expert-annotated route structures 161
Expert-scored step occurrences 519
Distinct reactions in the expert-scored cohorts 517
Individual expert step ratings 991
Consensus step scores 319
Total expert step-score entries 1,310
Original expert whole-route ratings 200
Model feedback annotations 15,945

Human versus LLM step scores

Rubric / assessor Full distinct reactions Step-score entries Flat distinct reactions / rows
Categorical rubric β€” human 202 991 202 / 991
Categorical rubric β€” LLMs, four models 202 12,808 202 / 12,808
Development rubric β€” human consensus 319 319 120 / 120

Four reactions overlap between the human cohorts: their full union is 517 reactions, not 521. The flat human union is 322 reactions with 1,111 score entries. The LLM reaction set is the same 202 reactions as the original human cohort.

Repeated ratings are separate score entries, not separate reactions. The enriched validation table represents the same 200 whole-route rating observations and must not be added to that total again.

Scores grouped by reaction

The default views contain one row per reaction with an assessments list. Grouped expert data has 322 reaction rows and 1,111 assessments. Each object keeps its score, feedback, rubric, evaluator and route reference together. Repeated equal ratings remain separate.

All seven grouped files use the same structure. All seven flat files contain one assessment per row, with the same fields plus reaction_smiles. These are alternative views of the same observations.

Reaction scores, feedback and scope

Every assessment includes score, local_feedback (the existing normalized category), local_feedback_text (the supplied explanation or null), rubric_id, assessor_type, assessor_id, model_id, confidence, and a stable assessment_id. Repository-relative source references identify the original record and reaction occurrence. See the field guide.

These are reaction-linked judgments made within a route: assessment_scope is step_in_route. Neither flattening nor restoring feedback turns them into intrinsic reaction-only feasibility scores.

  • Individual humans: 991 ratings; 364 have explanations, 627 are null.
  • Four LLMs: 12,808 ratings, all with explanations.
  • Included development consensus: 120 scores, no verified step-specific categories or explanations. Route-level reasons remain in the linked full records; they are not copied into step feedback.

A rubric describes what a score means. It does not supply a missing reaction-specific rationale. No new scores or generated explanations were added. The rubric IDs are expert_individual_and_llm and expert_development_consensus, matching sections in rubrics.json.

Schema 3 replaces the grouped scores / rubric_ids arrays with complete assessments objects. To obtain a score list, use [a["score"] for a in row["assessments"].

Scoring rubrics and structured rubric definitions cover the original expert questionnaire, the LLM judge and development consensus criteria verbatim, with source attribution and the complete original LLM prompt. Curation notes are labeled separately. LLM numeric scores are category-derived points, not confidence.

Of the 319 development scores, 120 map safely because every step in each included route has the same score. The remaining 199 have unresolved reaction ordering and remain in the full dataset. The complete expert corpus still has 1,310 score entries; the flat expert file is the safely assignable subset. Uniform scores resolve step ordering only; the development assessment-to-structure links remain inferred from unique canonical targets and matching step counts.

Files

File Content
all_routes.json Combined structural catalog
routes.json Original 50-route catalog
expert_feedback.json Individual expert reaction and route assessments
expert_consensus_development.json 121 development assessments, 319 step scores and 112 available structures
validation_enriched.json 50-route validation table with 16 anonymized raters, aggregates and model predictions
example_routes.json Ten example structures and route features
claude_opus48.json Claude evaluations and response text
gemini_35_pro.json Gemini evaluations; recorded model ID identifies Gemini 3.1 Pro
gpt55.json GPT evaluations and response text
llama31_70b.json Llama evaluations and response text
analysis_report.json Original-cohort descriptive statistics
cohort_analysis.json Cohort coverage, overlap and data-quality statistics

See analysis for coverage and label distributions.

Interpretation

  • Development assessment 4 and original route 34 share a complete structure.
  • Nine development assessments have no available route tree; their route link is null.
  • Mixed-score development routes retain recorded step order without an established mapping to reaction nodes. Only order-invariant consensus cases enter the flat export.
  • Development mean aggregation floors the mean; validation rounded consensus follows the observed half-up convention. Majority-vote ties remain explicit.
  • Missing observations are null. Bibliographic metadata and external URL spans have been removed; structures, scores, labels and analytical distinctions are retained.

Reading a file

After downloading the files, run from this repository's root:

import json
from pathlib import Path

dataset = json.loads(Path("clean/gpt55.json").read_text())
routes = {route["route_id"]: route for route in dataset["routes"]}
evaluation = dataset["records"][0]
route = routes[evaluation["route_id"]]

Read each nested file separately. The table viewer is configured only for the grouped subsets; observations is a loading label, not a train/test split. Verify the release with shasum -a 256 -c SHA256SUMS.

Citation

Zenodo record.

Downloads last month
3,052