release v0.2

README.md

@@ -13,7 +13,7 @@ tags:
   - cheminformatics
   - open-data
 size_categories:
-  - 10K<n<100K
+  - 100K<n<1M
 pretty_name: IRexp
 configs:
   - config_name: commercial
@@ -28,51 +28,74 @@ configs:
     data_files: data/irexp_non_commercial.jsonl.gz
   - config_name: other
     data_files: data/irexp_other.jsonl.gz
+  - config_name: structure_quarantine
+    data_files: data/structure_quarantine.jsonl.gz
 ---
 
-# IRexp — experimental IR band lists
+# IRexp: experimental IR band lists
 
-IRexp is an openly redistributable collection of **experimental infrared band lists** mined from open-access chemistry literature, often with co-reported ¹H/¹³C shift lists and resolved structures (SMILES / InChIKey).
+IRexp is an openly redistributable collection of **experimental infrared band lists** mined from open-access chemistry literature (PMC Open Access main text and supporting information) and a Chemotion deposit, often with co-reported ¹H/¹³C NMR strings and resolved structures (SMILES / InChIKey / SELFIES).
 
-> **Important:** IRexp contains **band lists** (peak positions in cm⁻¹), not digitised absorbance traces. This is the form reported in publication text and is not directly comparable to SDBS or NIST full spectra.
+> **Important:** IRexp contains **band lists** (peak positions in cm⁻¹), not digitised absorbance traces.
 
-This Hugging Face repository is the Scientific Data **dataset of record (DoR)** for the commercial (CC-BY / CC0) pool, plus separately licensed companions on the same revision. It is a **data release only** — no model accuracy tables, leaderboards, or benchmark scores are shipped here.
+This repository is the Scientific Data **dataset of record (DoR)** for the commercial (CC-BY / CC0) pool, plus separately licensed companions on the same revision. It is a data release only: no model scores are shipped here.
 
-## Files on this Hub revision
+## Files on this revision
 
 | Config / file | Records | Role |
 |---|---:|---|
-| `commercial` / `data/irexp_commercial.jsonl.gz` | **88,519** | **Dataset of record** — CC-BY + CC0; CC-BY-4.0 packaging |
-| `resolved_commercial` / `data/irexp_resolved_commercial.jsonl.gz` | 29,255 | Structure-linked commercial subset |
-| `train_no_bench_commercial` / `data/train_no_bench_commercial.jsonl.gz` | 29,109 | Commercial subset with held-out benchmark InChIKey-14 identities |
-| `sharealike` / `data/irexp_sharealike.jsonl.gz` | 1,897 | **Companion** — Chemotion + rare PMC ShareAlike; CC-BY-SA-4.0; **excluded from** the CC-BY DoR |
-| `non_commercial` / `data/irexp_non_commercial.jsonl.gz` | 21,823 | **Companion** — NC* (per-record policy licences, including CC-BY-NC-ND); **excluded from** the CC-BY DoR |
-| `other` / `data/irexp_other.jsonl.gz` | 5 | **Companion** — CC-BY-ND; **excluded from** the CC-BY DoR |
+| `commercial` / `data/irexp_commercial.jsonl.gz` | **163,931** | **Dataset of record**: CC-BY + CC0; CC-BY-4.0 packaging |
+| `resolved_commercial` / `data/irexp_resolved_commercial.jsonl.gz` | 67,235 | Commercial rows with a structure (tiers A–D) |
+| `train_no_bench_commercial` / `data/train_no_bench_commercial.jsonl.gz` | 67,093 | That subset without the 248 benchmark InChIKey-14 keys |
+| `sharealike` / `data/irexp_sharealike.jsonl.gz` | 1,897 | Companion: Chemotion + rare PMC ShareAlike; CC-BY-SA-4.0; **excluded from** the CC-BY DoR |
+| `non_commercial` / `data/irexp_non_commercial.jsonl.gz` | 65,007 | Companion: NC* (CC-BY-NC, CC-BY-NC-ND, CC-BY-NC-SA, stamped per row); **excluded from** the CC-BY DoR |
+| `other` / `data/irexp_other.jsonl.gz` | 5 | Companion: CC-BY-ND; **excluded from** the CC-BY DoR |
+| `structure_quarantine` / `data/structure_quarantine.jsonl.gz` | 8,431 | Ids only: tier-X structures withheld (8,132) plus formula–NMR diagnostic flags (299) |
 
-Also present: `data/f1_commercial_build_stats.json`, `NOTICE`, `LICENCE_REMEDIATION.md`, `scripts/release_check.py`.
+Also present: `data/seen_papers.txt.gz` (220,159 PMC identifiers scanned), `data/irexp_release_build_stats.json`, `data/SHA256SUMS.txt`, `NOTICE`, `LICENCE_REMEDIATION.md`, `scripts/release_check.py`, `scripts/build_release.py`.
 
-**Release note (this revision).** The commercial DoR was 88,545 on the prior Hub revision. After correcting thousands-separators in band lists, the ≥3-band inclusion rule was reapplied (−26 rows). Duplicate integer band positions within a list were removed in place (row count unchanged). The structure-linked subset was rebuilt as commercial records with a non-empty SMILES string (29,255). Six commercial rows with non-empty SMILES had `has_structure=false`; flags were set to `true` to match Table 1 (True when SMILES present). The five CC-BY-ND rows are published as the `other` companion for licence-pool symmetry with published NC-ND; empty/unknown (8,963) remain off this Hub revision.
+Public downloadable total (commercial + non-commercial + ShareAlike + ND): **230,840**. The research corpus is **239,735** records; the 8,895 rows with an empty or unknown licence are not redistributed.
 
-**Not on this Hub revision:** the multi-licence research total (`irexp.jsonl.gz`, n=121,233), empty/unknown (8,963), harvest snapshots, and quarantine lists. Those remain research-corpus material.
+## What changed in this release
 
-The research corpus (construction / licence-pool accounting) is 121,233 records (119,345 PMC + 1,888 Chemotion). Only the commercial pool is the public DoR.
+1. **IR-attachment fix (v4 parser).** IR blocks are assigned within each compound's own section of the experimental text instead of to the nearest NMR block. Rows whose IR belonged to a neighbouring compound were repaired or removed. The PMC main-text part is now 116,320 research rows (previously 119,345).
+2. **Supporting-information expansion.** 121,527 new band lists from the SI files of 5,688 PMC Open Access papers, parsed with the same rules (≥3 bands, 400–4000 cm⁻¹, within-list integer-band dedup, dedup against existing rows, licence join per article). Rows from articles without a redistributable licence were not kept.
+3. **Strict structure tiers.** Every structure now carries `structure_tier`:
 
-## Licence
+| Tier | Rule | Research corpus (239,735) |
+|---|---|---:|
+| A | formula/HRMS match in the compound's own block **plus** an independent second method (PubChem name lookup or OCSR drawing) agreeing on InChIKey-14 | 7,366 |
+| B | formula/HRMS match in the own block | 48,669 |
+| C | no formula; the source passed a ≥97% held-out agreement check | 21,455 |
+| D | IR-only rows, OCSR without a formula, or the Chemotion deposit structure on IR-only rows | 24,134 |
+| X | contradiction (for example IR and NMR from two different compounds, invalid or salt structure): **structure withheld**, row kept | 8,473 |
+| none | no structure | 129,638 |
+
+Tiers A–B: 56,035 (23.4%). Tiers A–D (rows with a released structure): 101,624 (42.4%). OCSR alone is never tier A. Methods ship only at ≥97% held-out precision.
 
-- Commercial DoR: Hub YAML `cc-by-4.0` packaging; per-record source licences stamped.
-- ShareAlike companion: source licence CC-BY-SA-4.0 — do not relicense as CC-BY.
-- Non-commercial companion: Hub YAML also lists `cc-by-nc-4.0` so the card reflects the NC pool; **policy licences on each NC row still apply** (CC-BY-NC, CC-BY-NC-ND, CC-BY-NC-SA as stamped). Honour NC terms; not part of the CC-BY DoR.
-- Other companion: five CC-BY-ND rows (`license_pool=other`); not part of the CC-BY DoR.
-- Every row carries `license`, `license_pool`, `license_raw`, `license_source` (plus source DOI / PMCID where available).
+## New and changed columns
 
-## Quality flags (commercial DoR)
+- `structure_tier` (A/B/C/D/X/null), `structure_method`, `structure_second_method` (tier A only), `formula_match`, `structure_flags` (`formula_conflict`, `pubchem_conflict`, IR plausibility flags; flag-only).
+- `source_set`: `pmc_v4` (main text), `si_2026` (supporting information), `chemotion`.
+- SI rows also carry `si_label`, `si_name`, `ir_raw` (author IR string) and `parser`.
+- `ir_bands_old_cm-1` on SI rows is the list before within-list integer dedup. `ir_table_flatten_suspect` is null on SI rows (not computed for that set).
 
-Flag-only fields; flagged rows are retained in the commercial n=88,519:
+## Quality checks (automated and AI audits; not human audits)
 
-- `ir_shared_in_paper` — identical IR list shared by ≥2 records in one paper.
-- `ir_table_flatten_suspect` — suspected table-flatten / column-misread list.
+- SI band lists, AI audit of 150 random rows: bands correct 149/150, compound attribution 149/150, experimental IR 150/150.
+- SI structures before strict tiers, AI audit of 150 random resolved rows: 146/150 correct (Wilson 95% CI 93.3–99.0%).
+- Formula-check promotions on the main text, AI audit: 40/40 correct (Wilson 91.2–100%).
+- OCSR label-B rule on 100 held-out papers: 104/104 (Wilson 96.4–100%). On the full main-text OCSR run the same rule scored 95.1%, below the 97% bar, so those structures were shipped as tier D only.
+- IR/NMR split headings: 12/12 verified as genuine separate compounds (these rows are tier X).
+
+## Licence
 
-Filter locally if needed. Audit fields `ir_bands_old_cm-1`, `ir_refetch_confirmed`, and `ir_refetch_merged` document reparse and refetch history.
+- Commercial DoR: YAML `cc-by-4.0` packaging; per-record source licences stamped.
+- ShareAlike companion: CC-BY-SA-4.0. Do not relicense as CC-BY.
+- Non-commercial companion: YAML lists `cc-by-nc-4.0` so the card reflects the NC pool; **the per-row policy licences apply** (CC-BY-NC, CC-BY-NC-ND, CC-BY-NC-SA). Not part of the CC-BY DoR.
+- Other companion: five CC-BY-ND rows; not part of the CC-BY DoR.
+- Every row carries `license`, `license_pool`, `license_raw`, `license_source`, plus `source_doi` / `pmcid` / `doi`.
+- Only extracted numbers and identifiers are redistributed; no article text, figures or PDFs.
 
 ## Load
 
@@ -80,14 +103,17 @@ Filter locally if needed. Audit fields `ir_bands_old_cm-1`, `ir_refetch_confirme
 from datasets import load_dataset
 
 dor = load_dataset("ilkhamfy/IRexp", "commercial", split="train")
-sa = load_dataset("ilkhamfy/IRexp", "sharealike", split="train")
+res = dor.filter(lambda r: r["structure_tier"] in ("A", "B"))  # formula-verified structures
 nc = load_dataset("ilkhamfy/IRexp", "non_commercial", split="train")
-other = load_dataset("ilkhamfy/IRexp", "other", split="train")  # CC-BY-ND (n=5)
 ```
 
+## Training without benchmark leakage
+
+`train_no_bench_commercial` withholds the 248 InChIKey-14 keys of the locked benchmark rounds (locked 194 and electrolyte). Keys of later expansion rounds are not filtered, so fine-tuned models should be scored on the locked 194.
+
 ## Citation
 
-Cite the Scientific Data Data Descriptor (when published) and this Hub revision. Zenodo archival DOI for the commercial primary file: https://doi.org/10.5281/zenodo.22822285.
+Cite the Scientific Data Data Descriptor (when published) and this Hub revision. Zenodo version DOI for the commercial files of this release: https://doi.org/10.5281/zenodo.23463965 (concept DOI, always the latest version: https://doi.org/10.5281/zenodo.22822284).
 
 ## Links
 
@@ -95,3 +121,7 @@ Cite the Scientific Data Data Descriptor (when published) and this Hub revision.
 - Manuscript / figures: https://github.com/IlkhamFY/IRexp
 - Extraction tooling: https://github.com/IlkhamFY/spectro-agent
 - Licence details: `NOTICE`, `LICENCE_REMEDIATION.md`
+
+## Previous revision
+
+`4312254c279300f881ad12671476e877027ed3da`: commercial DoR 88,519 (archived at https://doi.org/10.5281/zenodo.23023528).

data/SHA256SUMS.txt

@@ -1,8 +1,9 @@
-a9403e7b6f6c1d6181f40fd78d66a7c3f202cc15d363bcb25869db5ed59c8eb5  f1_commercial_build_stats.json
-d1e363fafe7d29985eb80bf17e1c218ad826268ee4cfa76ab5895be3da011798  irexp_commercial.jsonl.gz
-8c60800c9222eb48f2e1256c13380ab6a7c4680f7bac515f45a2373d0a3e8c8b  irexp_non_commercial.jsonl.gz
-b8710a52aa97537f708e358751783a1ea8ab927eb24b68ef584326276aaa6def  irexp_other.jsonl.gz
-145adcc29b597774f4d0134e5dc42cdc4acc459a906f6e406b69c3e5fb6232a7  irexp_resolved_commercial.jsonl.gz
-21af64a91fb51e0694a3f4c665b89349de86ce3826197ecd6f6357fb26501e6e  irexp_sharealike.jsonl.gz
-b7c05953558c0e6b48655e01792294823d866c58ee64a5cc32ae626c50cc531d  patch_meta_presubmit.json
-509fbd9eaff073f0c313424df62ce0794ee76c8c014fbb3d7860e96e544cc418  train_no_bench_commercial.jsonl.gz
+3c784a3b1b5aba753127c25ff256097bba6e3dd17e864af6671b6d68ccd66b2f  irexp_commercial.jsonl.gz
+6d777812ce6625880040998bfd43fc421a4f215ba7e46519ebf478625da8b2be  irexp_resolved_commercial.jsonl.gz
+00069af9bb4b3e5f5615d3db605d72bb6b837e78127deb6aed643d9cfc80e52d  train_no_bench_commercial.jsonl.gz
+6dfb27544a3b87feb2909e66614c884fde0c9ad61046eff9999c048a09fc220c  irexp_non_commercial.jsonl.gz
+494d306dce39ced51089a3a3728054f993b220d282c77ce115c5305beea73596  irexp_sharealike.jsonl.gz
+52e4a0a6ba1ba7136896df0015812973bffdf4b90c855af98a642b767cebcf25  irexp_other.jsonl.gz
+29808d9d7a9ff8cd8be25295db3d5eed3e95272ce945331784314133c2534f51  structure_quarantine.jsonl.gz
+ad0c91414f8a85463ae847a688ece35f5b5b4c87cb47e5896027114fd6587282  seen_papers.txt.gz
+b4930118e4b8cae62705b197694a61dc9db65687565e4e7aef513c78001211da  irexp_release_build_stats.json

data/f1_commercial_build_stats.json

@@ -1,51 +0,0 @@
-{
-  "commercial": {
-    "label": "commercial",
-    "n_commercial": 88519,
-    "n_previous_hub": 88545,
-    "dropped_lt3": 26,
-    "deduped_integer_bands": 1432,
-    "note": "Reapplied \u22653-band filter after thousands-separator reparse; order-preserving unique of integer cm\u22121 within each list. ND (5) published separately as other companion (not merged into commercial).",
-    "bytes_approx": 10853218
-  },
-  "resolved_commercial": {
-    "label": "resolved_commercial",
-    "n_commercial": 29255,
-    "n_previous_hub": 28899,
-    "invariant": "resolved_commercial = commercial \u2229 nonempty(smiles)",
-    "bytes_approx": 4646831
-  },
-  "train_no_bench_commercial": {
-    "label": "train_no_bench_commercial",
-    "n_commercial": 29109,
-    "n_previous_hub": 28753,
-    "skipped_benchmark_ids": 146,
-    "invariant": "train_no_bench_commercial \u2286 resolved_commercial",
-    "bytes_approx": 4626653
-  },
-  "non_commercial": {
-    "n": 21823,
-    "deduped_integer_bands": 319
-  },
-  "sharealike": {
-    "n": 1897
-  },
-  "patched_from_hub_rev": "50d615533563e5071efd27bd7a0f93aa1025b69b",
-  "has_structure_presubmit": {
-    "flipped_to_true": 6,
-    "ids": [
-      "ABWOQVYUYOZDPM-UHFFFAOYSA-N",
-      "HEKOVRRPURYUEM-UHFFFAOYSA-N",
-      "MFQSCMXWQUQCBL-UHFFFAOYSA-N",
-      "NWIJCILLNNWXQP-UHFFFAOYSA-N",
-      "SSVMNDFIYNDYTJ-UHFFFAOYSA-N",
-      "VFSNNVLONLFMCD-UHFFFAOYSA-N"
-    ],
-    "note": "Aligned with Table 1: has_structure True when SMILES present."
-  },
-  "other_nd": {
-    "n": 5,
-    "published_on_hub": true,
-    "file": "data/irexp_other.jsonl.gz"
-  }
-}

data/irexp_commercial.jsonl.gz

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d1e363fafe7d29985eb80bf17e1c218ad826268ee4cfa76ab5895be3da011798
-size 10853220
+oid sha256:3c784a3b1b5aba753127c25ff256097bba6e3dd17e864af6671b6d68ccd66b2f
+size 23256461

data/irexp_non_commercial.jsonl.gz

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8c60800c9222eb48f2e1256c13380ab6a7c4680f7bac515f45a2373d0a3e8c8b
-size 2419718
+oid sha256:6dfb27544a3b87feb2909e66614c884fde0c9ad61046eff9999c048a09fc220c
+size 9055476

data/irexp_other.jsonl.gz

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b8710a52aa97537f708e358751783a1ea8ab927eb24b68ef584326276aaa6def
-size 700
+oid sha256:52e4a0a6ba1ba7136896df0015812973bffdf4b90c855af98a642b767cebcf25
+size 823

data/irexp_release_build_stats.json

@@ -0,0 +1,198 @@
+{
+ "counts": {
+  "commercial": 163931,
+  "resolved_commercial": 67235,
+  "train_no_bench_commercial": 67093,
+  "non_commercial": 65007,
+  "sharealike": 1897,
+  "other": 5,
+  "structure_quarantine": 8431
+ },
+ "files": {
+  "commercial": {
+   "records": 163931,
+   "structure_linked": 67235,
+   "unique_inchikey": 64107,
+   "bands": 1916743,
+   "papers": 15390,
+   "tiers": {
+    "none": 90526,
+    "B": 31800,
+    "C": 15791,
+    "X": 6170,
+    "D": 15146,
+    "A": 4498
+   },
+   "by_set": {
+    "pmc_v4": 85878,
+    "si_2026": 78053
+   },
+   "full_quadruple": 36927,
+   "lt3": 0,
+   "band_out_of_range": 0
+  },
+  "non_commercial": {
+   "records": 65007,
+   "structure_linked": 27917,
+   "unique_inchikey": 27170,
+   "bands": 780919,
+   "papers": 4451,
+   "tiers": {
+    "B": 13986,
+    "none": 35214,
+    "C": 4871,
+    "D": 6507,
+    "X": 1876,
+    "A": 2553
+   },
+   "by_set": {
+    "pmc_v4": 21533,
+    "si_2026": 43474
+   },
+   "full_quadruple": 13739,
+   "lt3": 0,
+   "band_out_of_range": 0
+  },
+  "sharealike": {
+   "records": 1897,
+   "structure_linked": 1804,
+   "unique_inchikey": 1804,
+   "bands": 72717,
+   "papers": 2,
+   "tiers": {
+    "none": 7,
+    "B": 2,
+    "D": 1802,
+    "X": 86
+   },
+   "by_set": {
+    "pmc_v4": 9,
+    "chemotion": 1888
+   },
+   "full_quadruple": 2,
+   "lt3": 0,
+   "band_out_of_range": 0
+  },
+  "other": {
+   "records": 5,
+   "structure_linked": 0,
+   "unique_inchikey": 0,
+   "bands": 100,
+   "papers": 2,
+   "tiers": {
+    "none": 5
+   },
+   "by_set": {
+    "pmc_v4": 5
+   },
+   "full_quadruple": 0,
+   "lt3": 0,
+   "band_out_of_range": 0
+  }
+ },
+ "resolved_commercial": {
+  "records": 67235,
+  "structure_linked": 67235,
+  "unique_inchikey": 64107,
+  "bands": 758501,
+  "papers": 7022,
+  "tiers": {
+   "B": 31800,
+   "C": 15791,
+   "D": 15146,
+   "A": 4498
+  },
+  "by_set": {
+   "pmc_v4": 37116,
+   "si_2026": 30119
+  },
+  "full_quadruple": 36927,
+  "lt3": 0,
+  "band_out_of_range": 0
+ },
+ "train_no_bench_commercial": {
+  "records": 67093,
+  "structure_linked": 67093,
+  "unique_inchikey": 63971,
+  "bands": 756907,
+  "papers": 7019,
+  "tiers": {
+   "B": 31710,
+   "C": 15749,
+   "D": 15146,
+   "A": 4488
+  },
+  "by_set": {
+   "pmc_v4": 36974,
+   "si_2026": 30119
+  },
+  "full_quadruple": 36786,
+  "lt3": 0,
+  "band_out_of_range": 0
+ },
+ "released_total": 230840,
+ "research_corpus": {
+  "n": 239735,
+  "tiers": {
+   "none": 129638,
+   "B": 48669,
+   "C": 21455,
+   "X": 8473,
+   "D": 24134,
+   "A": 7366
+  },
+  "AB": 56035,
+  "AB_pct": 23.37,
+  "AD": 101624,
+  "AD_pct": 42.39
+ },
+ "released_tiers": {
+  "none": 125752,
+  "B": 45788,
+  "C": 20662,
+  "X": 8132,
+  "D": 23455,
+  "A": 7051
+ },
+ "tiers_by_pool": {
+  "commercial/A": 4498,
+  "commercial/B": 31800,
+  "commercial/C": 15791,
+  "commercial/D": 15146,
+  "commercial/X": 6170,
+  "commercial/none": 90526,
+  "non_commercial/A": 2553,
+  "non_commercial/B": 13986,
+  "non_commercial/C": 4871,
+  "non_commercial/D": 6507,
+  "non_commercial/X": 1876,
+  "non_commercial/none": 35214,
+  "other/none": 5,
+  "sharealike/B": 2,
+  "sharealike/D": 1802,
+  "sharealike/X": 86,
+  "sharealike/none": 7
+ },
+ "stats": {
+  "v4_x_structure_removed_commercial": 4278,
+  "v4_structure_added_commercial": 2310,
+  "v4_structure_added_non_commercial": 533,
+  "v4_x_structure_removed_non_commercial": 815,
+  "v4_x_structure_removed_sharealike": 86,
+  "v4_rows": 109313,
+  "si_dup_int_band_rows_commercial": 537,
+  "si_dup_int_band_rows_non_commercial": 349,
+  "si_rows_in": 121527,
+  "si_rows_out": 121527
+ },
+ "seen_papers": {
+  "v4": 188016,
+  "total": 220159
+ },
+ "quarantine_reasons": {
+  "tier_X": 8132,
+  "c13_peaks_gt_carbons": 163,
+  "h1_integration_gt_formula_plus_2": 136
+ },
+ "benchmark_ik14": 248
+}

data/irexp_resolved_commercial.jsonl.gz

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:145adcc29b597774f4d0134e5dc42cdc4acc459a906f6e406b69c3e5fb6232a7
-size 4646811
+oid sha256:6d777812ce6625880040998bfd43fc421a4f215ba7e46519ebf478625da8b2be
+size 11633640

data/irexp_sharealike.jsonl.gz

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:21af64a91fb51e0694a3f4c665b89349de86ce3826197ecd6f6357fb26501e6e
-size 258270
+oid sha256:494d306dce39ced51089a3a3728054f993b220d282c77ce115c5305beea73596
+size 264884

data/patch_meta.json

@@ -1,240 +0,0 @@
-{
-  "commercial_old": 88545,
-  "commercial_new": 88519,
-  "dropped": [
-    {
-      "id": "82543b9311ceb3f6216d",
-      "pmcid": "PMC6271600",
-      "n_bands": 1,
-      "bands": [
-        471.5
-      ],
-      "in_drop_list": true
-    },
-    {
-      "id": "CRTKEGAHXINQAO-MKMNVTDBSA-N",
-      "pmcid": "PMC6271600",
-      "n_bands": 0,
-      "bands": [],
-      "in_drop_list": true
-    },
-    {
-      "id": "UFFZJOLWCBNFJJ-UHFFFAOYSA-N",
-      "pmcid": "PMC6271600",
-      "n_bands": 0,
-      "bands": [],
-      "in_drop_list": true
-    },
-    {
-      "id": "WLNOTYKLRHLKIX-MKMNVTDBSA-N",
-      "pmcid": "PMC6271600",
-      "n_bands": 1,
-      "bands": [
-        3811.0
-      ],
-      "in_drop_list": true
-    },
-    {
-      "id": "VVBDJRNCALMXCU-UHFFFAOYSA-N",
-      "pmcid": "PMC6271600",
-      "n_bands": 0,
-      "bands": [],
-      "in_drop_list": true
-    },
-    {
-      "id": "HFOOPPBDVOOJGT-UHFFFAOYSA-N",
-      "pmcid": "PMC6271600",
-      "n_bands": 1,
-      "bands": [
-        520.43
-      ],
-      "in_drop_list": true
-    },
-    {
-      "id": "2918cf238a7345c9990e",
-      "pmcid": "PMC6271600",
-      "n_bands": 2,
-      "bands": [
-        484.4,
-        548.0
-      ],
-      "in_drop_list": true
-    },
-    {
-      "id": "6be8723d27f7aead89e7",
-      "pmcid": "PMC11701777",
-      "n_bands": 0,
-      "bands": [],
-      "in_drop_list": true
-    },
-    {
-      "id": "BYSCDXUVBHRSJZ-UHFFFAOYSA-N",
-      "pmcid": "PMC11625556",
-      "n_bands": 2,
-      "bands": [
-        2908.0,
-        1725.0
-      ],
-      "in_drop_list": true
-    },
-    {
-      "id": "KNQPSCBYHORSRO-UHFFFAOYSA-N",
-      "pmcid": "PMC10552108",
-      "n_bands": 2,
-      "bands": [
-        1655.0,
-        1485.0
-      ],
-      "in_drop_list": true
-    },
-    {
-      "id": "LANOVFBPKPZBSX-UHFFFAOYSA-N",
-      "pmcid": "PMC10552108",
-      "n_bands": 2,
-      "bands": [
-        1669.0,
-        1303.0
-      ],
-      "in_drop_list": true
-    },
-    {
-      "id": "a495b878d51ec42a6ad6",
-      "pmcid": "PMC10833406",
-      "n_bands": 0,
-      "bands": [],
-      "in_drop_list": true
-    },
-    {
-      "id": "e280d5487f1c8913cb3d",
-      "pmcid": "PMC8819441",
-      "n_bands": 0,
-      "bands": [],
-      "in_drop_list": true
-    },
-    {
-      "id": "6b8bc59f4b554ea17b32",
-      "pmcid": "PMC9010074",
-      "n_bands": 1,
-      "bands": [
-        2315.0
-      ],
-      "in_drop_list": true
-    },
-    {
-      "id": "b84dd30e882130a2d1ef",
-      "pmcid": "PMC9688225",
-      "n_bands": 2,
-      "bands": [
-        2700.0,
-        1709.33
-      ],
-      "in_drop_list": true
-    },
-    {
-      "id": "0406f3ebb3d9753f1ce5",
-      "pmcid": "PMC8303174",
-      "n_bands": 0,
-      "bands": [],
-      "in_drop_list": true
-    },
-    {
-      "id": "245d841904ba9d33a218",
-      "pmcid": "PMC8707475",
-      "n_bands": 0,
-      "bands": [],
-      "in_drop_list": true
-    },
-    {
-      "id": "513ba0e6d6a9227f0df8",
-      "pmcid": "PMC7054417",
-      "n_bands": 0,
-      "bands": [],
-      "in_drop_list": true
-    },
-    {
-      "id": "ad98381d8f119ca8d637",
-      "pmcid": "PMC6727057",
-      "n_bands": 0,
-      "bands": [],
-      "in_drop_list": true
-    },
-    {
-      "id": "8abe447b7a282f6be2f2",
-      "pmcid": "PMC6099295",
-      "n_bands": 2,
-      "bands": [
-        819.0,
-        428.0
-      ],
-      "in_drop_list": true
-    },
-    {
-      "id": "be20e932e87ac1a73e55",
-      "pmcid": "PMC6099295",
-      "n_bands": 1,
-      "bands": [
-        400.0
-      ],
-      "in_drop_list": true
-    },
-    {
-      "id": "a1b91bf246ef651d0f90",
-      "pmcid": "PMC6099295",
-      "n_bands": 1,
-      "bands": [
-        445.0
-      ],
-      "in_drop_list": true
-    },
-    {
-      "id": "e5d62449112ccad5f58e",
-      "pmcid": "PMC5209758",
-      "n_bands": 0,
-      "bands": [],
-      "in_drop_list": true
-    },
-    {
-      "id": "0228d5168d65c201c5d7",
-      "pmcid": "PMC4598625",
-      "n_bands": 2,
-      "bands": [
-        420.0,
-        3700.0
-      ],
-      "in_drop_list": true
-    },
-    {
-      "id": "5a2ddff7c25ccc89770f",
-      "pmcid": "PMC4174874",
-      "n_bands": 0,
-      "bands": [],
-      "in_drop_list": true
-    },
-    {
-      "id": "181bf87501cfe43138fd",
-      "pmcid": "PMC4257448",
-      "n_bands": 0,
-      "bands": [],
-      "in_drop_list": true
-    }
-  ],
-  "deduped_commercial": 1432,
-  "resolved_old": 28899,
-  "resolved_new": 29255,
-  "tnb_old": 28753,
-  "tnb_new": 29109,
-  "tnb_skip_id": 146,
-  "tnb_skip_extra_ik14": 0,
-  "held_ik14_n": 142,
-  "deduped_nc": 319,
-  "nc_n": 21823,
-  "sa_n": 1897,
-  "hs_false_with_smiles_in_comm": [
-    "HEKOVRRPURYUEM-UHFFFAOYSA-N",
-    "MFQSCMXWQUQCBL-UHFFFAOYSA-N",
-    "NWIJCILLNNWXQP-UHFFFAOYSA-N",
-    "VFSNNVLONLFMCD-UHFFFAOYSA-N",
-    "ABWOQVYUYOZDPM-UHFFFAOYSA-N",
-    "SSVMNDFIYNDYTJ-UHFFFAOYSA-N"
-  ]
-}

data/patch_meta_presubmit.json

@@ -1,30 +0,0 @@
-{
-  "base_hub_rev": "fc238e3b600f556045bc9a3f7a37738734620cd8",
-  "fixes": {
-    "has_structure_flip_ids": [
-      "ABWOQVYUYOZDPM-UHFFFAOYSA-N",
-      "HEKOVRRPURYUEM-UHFFFAOYSA-N",
-      "MFQSCMXWQUQCBL-UHFFFAOYSA-N",
-      "NWIJCILLNNWXQP-UHFFFAOYSA-N",
-      "SSVMNDFIYNDYTJ-UHFFFAOYSA-N",
-      "VFSNNVLONLFMCD-UHFFFAOYSA-N"
-    ],
-    "decision": "flip has_structure=true (Table 1: True when SMILES present); RDKit parses + InChIKey match",
-    "nd_publish": "add data/irexp_other.jsonl.gz (5 CC-BY-ND) as Hub companion; commercial DoR N unchanged 88519"
-  },
-  "counts": {
-    "commercial": 88519,
-    "resolved_commercial": 29255,
-    "train_no_bench_commercial": 29109,
-    "sharealike": 1897,
-    "non_commercial": 21823,
-    "other_nd": 5
-  },
-  "public_bands": {
-    "commercial": 953310,
-    "non_commercial": 230596,
-    "sharealike": 72717,
-    "sum": 1256623,
-    "other_nd": 101
-  }
-}

data/seen_papers.txt.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ad0c91414f8a85463ae847a688ece35f5b5b4c87cb47e5896027114fd6587282
+size 586158

data/structure_quarantine.jsonl.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:29808d9d7a9ff8cd8be25295db3d5eed3e95272ce945331784314133c2534f51
+size 134274

data/train_no_bench_commercial.jsonl.gz

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:509fbd9eaff073f0c313424df62ce0794ee76c8c014fbb3d7860e96e544cc418
-size 4626635
+oid sha256:00069af9bb4b3e5f5615d3db605d72bb6b837e78127deb6aed643d9cfc80e52d
+size 11613493

scripts/build_release.py

@@ -0,0 +1,159 @@
+"""Build the combined IRexp release (v4 IR-attachment fix + 2026 SI expansion) with strict structure tiers.
+Inputs are the final local artifacts; nothing is fetched. Output: ../data (chmod 600)."""
+import gzip,json,os,re,collections,hashlib
+from rdkit import Chem, RDLogger; RDLogger.DisableLog('rdApp.*')
+import selfies as sf
+OUT='/workspace/irexp_release_237k/data/'
+V4R='/workspace/irexp_v4_build/release/data/'
+V4T='/workspace/irexp_growth/RESOLVE_237k/work/v4_tiers_final.jsonl.gz'
+SIH='/workspace/irexp_growth/harvest/out/'
+SIT='/workspace/irexp_growth/audit_rudra/new_tiers_final.jsonl'
+SIS='/workspace/irexp_growth/structures/'
+BENCH=set(json.load(open('/workspace/irexp_v4_build/work/bench_ik14.json')))
+LIC={'CC BY':'CC-BY','CC BY-NC':'CC-BY-NC','CC BY-NC-ND':'CC-BY-NC-ND','CC BY-NC-SA':'CC-BY-NC-SA','CC BY-SA':'CC-BY-SA','CC0':'CC0'}
+def mol_fields(smi):
+    m=Chem.MolFromSmiles(smi) if smi else None
+    if m is None: return None
+    can=Chem.MolToSmiles(m)
+    try: s=sf.encoder(can)
+    except Exception: s=None
+    return dict(smiles=can,inchikey=Chem.MolToInchiKey(m),selfies=s,_mol=m)
+import sys as _sys
+_sys.path.insert(0,'/workspace/irexp_v4_build/spectro-agent'); _sys.path.insert(0,'/workspace/irexp_v4_build/spectro-agent/scripts')
+from quarantine_structure_nmr import gate_record as _gate_record
+def nmr_check(m,h,c):
+    """spectro-agent scripts/quarantine_structure_nmr.py gate_record (same rules as the 2026-08-26 quarantine)."""
+    return [x for x in _gate_record({'smiles':Chem.MolToSmiles(m),'h_nmr':h,'c_nmr':c,'ir_bands_cm-1':[]}) if x!='no_smiles']
+def bands_ok(b):
+    return all(350<=float(x)<=4000 for x in b)
+pools=collections.defaultdict(list); quar=[]; stats=collections.Counter(); tiers=collections.Counter(); tiers_pool=collections.Counter()
+# ---------- v4 rows ----------
+T={}
+for l in gzip.open(V4T,'rt'):
+    d=json.loads(l); T[d['id']]=d
+for f,pool in [('irexp_commercial','commercial'),('irexp_non_commercial','non_commercial'),('irexp_sharealike','sharealike'),('irexp_other','other')]:
+    for l in gzip.open(V4R+f+'.jsonl.gz','rt'):
+        d=json.loads(l); t=T[d['id']]; tier=t['tier_final']
+        d['source_set']='pmc_v4' if pool!='sharealike' or d.get('pmcid') else 'chemotion'
+        if pool=='sharealike' and d.get('source'): d['source_set']='chemotion' if 'hemotion' in str(d.get('source')) else 'pmc_v4'
+        flags=[]
+        if t.get('formula_conflict'): flags.append('formula_conflict')
+        if t.get('pubchem_conflict'): flags.append('pubchem_conflict')
+        mf=None
+        if tier in 'ABCD' and t.get('smiles'):
+            mf=mol_fields(t['smiles'])
+            if mf is None: stats['v4_unparseable']+=1; tier='none'
+        if tier=='X':
+            if d.get('smiles'): stats['v4_x_structure_removed_'+pool]+=1
+            quar.append(dict(id=d['id'],license_pool=pool,source_set=d['source_set'],reasons=['tier_X:'+(t.get('x_reason') or t.get('base_why') or 'contradiction')]))
+        if mf:
+            if d.get('smiles') and Chem.MolToSmiles(Chem.MolFromSmiles(d['smiles']))==mf['smiles'] and d.get('selfies'):
+                mf['selfies']=d['selfies']
+            if not d.get('smiles'): stats['v4_structure_added_'+pool]+=1
+            d.update(smiles=mf['smiles'],inchikey=mf['inchikey'],selfies=mf['selfies'],has_structure=True)
+        else:
+            d.update(smiles=None,inchikey=None,selfies=None,has_structure=False)
+        d['structure_tier']=tier if tier in 'ABCD' else ('X' if tier=='X' else None)
+        d['structure_method']=(('ocsr:'+str(t['src_ocsr'].get('method'))) if t.get('src_ocsr') else t.get('method')) if mf else None
+        d['formula_match']=t.get('fm') if mf else None
+        d['structure_second_method']=('pubchem' if t.get('a_strict') else None) if tier=='A' else None
+        d['structure_flags']=flags
+        if mf:
+            q=nmr_check(mf['_mol'],d.get('h_nmr'),d.get('c_nmr'))
+            if q: quar.append(dict(id=d['id'],license_pool=pool,source_set=d['source_set'],inchikey=mf['inchikey'],reasons=q))
+        tiers[d['structure_tier'] or 'none']+=1; tiers_pool[(pool,d['structure_tier'] or 'none')]+=1
+        pools[pool].append(d)
+stats['v4_rows']=sum(len(v) for v in pools.values())
+# ---------- SI expansion ----------
+NT={}
+for l in open(SIT):
+    d=json.loads(l); NT[d['id']]=d
+SFL={}
+for f in ['structures_commercial.jsonl.gz','structures_non_commercial.jsonl.gz']:
+    for l in gzip.open(SIS+f,'rt'):
+        r=json.loads(l); SFL[r['id']]=r.get('ir_flags') or []
+nsi=0
+for p in ['commercial','non_commercial']:
+    for l in gzip.open(SIH+f'irexp_si_{p}.jsonl.gz','rt'):
+        s=json.loads(l); nsi+=1; t=NT[s['id']]; tier=t['tier']
+        b0=s['ir_bands_cm-1']; seenb=set(); nb=[]
+        for x in b0:
+            k=int(round(float(x)))
+            if k not in seenb: seenb.add(k); nb.append(x)
+        if len(nb)!=len(b0): stats['si_dup_int_band_rows_'+p]+=1
+        if len(nb)<3: stats['si_dropped_lt3_after_dedup_'+p]+=1; continue
+        lic=LIC.get(s['license'],s['license'])
+        d=dict(id=s['id'],**{'ir_bands_cm-1':nb,'ir_bands_old_cm-1':b0},ir_source=s['ir_source'],h_nmr=s.get('h_nmr'),c_nmr=s.get('c_nmr'),
+               smiles=None,selfies=None,inchikey=None,has_structure=False,source_doi='PMC:'+s['pmcid'][3:],license=lic,license_raw=s['license'],
+               license_pool=p,license_source='pmc_oa_s3',pmcid=s['pmcid'],doi=s.get('doi'),source_set='si_2026',si_label=s.get('label'),si_name=s.get('name'),
+               ir_raw=s.get('ir'),parser=s.get('parser'),ir_shared_in_paper=False,ir_table_flatten_suspect=None,ir_refetch_confirmed=False,ir_refetch_merged=False)
+        mf=None
+        if tier in ('A','B','C','D') and t.get('smiles'):
+            mf=mol_fields(t['smiles'])
+            if mf is None: stats['si_unparseable']+=1; tier='none'
+        if mf:
+            d.update(smiles=mf['smiles'],inchikey=mf['inchikey'],selfies=mf['selfies'],has_structure=True)
+            q=nmr_check(mf['_mol'],d.get('h_nmr'),d.get('c_nmr'))
+            if q: quar.append(dict(id=d['id'],license_pool=p,source_set='si_2026',inchikey=mf['inchikey'],reasons=q))
+        if tier=='X': quar.append(dict(id=d['id'],license_pool=p,source_set='si_2026',reasons=['tier_X:'+(t.get('x_reason') or 'contradiction')]))
+        d['structure_tier']=tier if tier in ('A','B','C','D','X') else None
+        meth=t.get('method')
+        if t.get('src')=='ocsr': meth='ocsr:'+str(meth)
+        d['structure_method']=meth if mf else None
+        d['formula_match']=t.get('formula_match') if mf else None
+        sm=[c for c in t.get('confirm',[]) if 'DISAGREES' not in c]
+        d['structure_second_method']=('; '.join(sm) or None) if d['structure_tier']=='A' else None
+        fl=list(SFL.get(s['id'],[]))
+        if any('DISAGREES' in c for c in t.get('confirm',[])): fl.append('pubchem_conflict')
+        d['structure_flags']=fl
+        tiers[d['structure_tier'] or 'none']+=1; tiers_pool[(p,d['structure_tier'] or 'none')]+=1
+        pools[p].append(d)
+stats['si_rows_in']=nsi
+grp=collections.Counter((r['pmcid'],tuple(int(round(float(x))) for x in r['ir_bands_cm-1'])) for p in ('commercial','non_commercial') for r in pools[p] if r['source_set']=='si_2026')
+for p in ('commercial','non_commercial'):
+    for r in pools[p]:
+        if r['source_set']=='si_2026' and grp[(r['pmcid'],tuple(int(round(float(x))) for x in r['ir_bands_cm-1']))]>1: r['ir_shared_in_paper']=True; stats['si_ir_shared_in_paper']+=1
+stats['si_rows_out']=sum(1 for p in ('commercial','non_commercial') for r in pools[p] if r['source_set']=='si_2026')
+# ---------- write ----------
+def w(name,rows):
+    path=OUT+name
+    with gzip.open(path,'wt',compresslevel=9) as g:
+        for r in rows: g.write(json.dumps(r,ensure_ascii=False)+'\n')
+    os.chmod(path,0o600); return len(rows)
+counts={}
+counts['commercial']=w('irexp_commercial.jsonl.gz',pools['commercial'])
+res=[r for r in pools['commercial'] if r['has_structure']]
+counts['resolved_commercial']=w('irexp_resolved_commercial.jsonl.gz',res)
+tnb=[r for r in res if r['inchikey'][:14] not in BENCH]
+counts['train_no_bench_commercial']=w('train_no_bench_commercial.jsonl.gz',tnb)
+counts['non_commercial']=w('irexp_non_commercial.jsonl.gz',pools['non_commercial'])
+counts['sharealike']=w('irexp_sharealike.jsonl.gz',pools['sharealike'])
+counts['other']=w('irexp_other.jsonl.gz',pools['other'])
+counts['structure_quarantine']=w('structure_quarantine.jsonl.gz',quar)
+# seen papers: v4 list + SI list
+seen=set(l.strip() for l in gzip.open('/workspace/irexp_v4_build/IRexp/data/seen_papers.txt.gz','rt') if l.strip())
+n4=len(seen); seen|=set(l.strip() for l in open('/workspace/irexp_growth/harvest/seen_papers_si.txt') if l.strip())
+with gzip.open(OUT+'seen_papers.txt.gz','wt') as g: g.write('\n'.join(sorted(seen))+'\n')
+os.chmod(OUT+'seen_papers.txt.gz',0o600)
+def agg(rows):
+    st=[r for r in rows if r['has_structure']]
+    return dict(records=len(rows),structure_linked=len(st),unique_inchikey=len({r['inchikey'] for r in st}),
+                bands=sum(len(r['ir_bands_cm-1']) for r in rows),papers=len({r.get('pmcid') or r.get('source_doi') for r in rows}),
+                tiers=dict(collections.Counter(r['structure_tier'] or 'none' for r in rows)),
+                by_set=dict(collections.Counter(r['source_set'] for r in rows)),
+                full_quadruple=sum(1 for r in st if r.get('h_nmr') and r.get('c_nmr')),
+                lt3=sum(1 for r in rows if len(r['ir_bands_cm-1'])<3),
+                band_out_of_range=sum(1 for r in rows if not bands_ok(r['ir_bands_cm-1'])))
+# research corpus (incl. empty/unknown licence pool, not redistributed)
+rc=collections.Counter(); 
+for l in gzip.open(V4T,'rt'): rc[json.loads(l)['tier_final']]+=1
+for d in NT.values(): rc[d['tier'] or 'none']+=1
+N=sum(rc.values())
+rep=dict(counts=counts,files={k:agg(v) for k,v in pools.items()},resolved_commercial=agg(res),train_no_bench_commercial=agg(tnb),
+         released_total=sum(len(v) for v in pools.values()),
+         research_corpus=dict(n=N,tiers=dict(rc),AB=rc['A']+rc['B'],AB_pct=round(100*(rc['A']+rc['B'])/N,2),AD=sum(rc[k] for k in 'ABCD'),AD_pct=round(100*sum(rc[k] for k in 'ABCD')/N,2)),
+         released_tiers=dict(tiers),tiers_by_pool={'%s/%s'%k:v for k,v in sorted(tiers_pool.items())},
+         stats=dict(stats),seen_papers=dict(v4=n4,total=len(seen)),quarantine_reasons=dict(collections.Counter(r['reasons'][0].split(':')[0] for r in quar)),
+         benchmark_ik14=len(BENCH))
+json.dump(rep,open(OUT+'irexp_release_build_stats.json','w'),indent=1); os.chmod(OUT+'irexp_release_build_stats.json',0o600)
+print(json.dumps(rep,indent=1))