Add README.md

README.md

@@ -0,0 +1,159 @@
+---
+language:
+  - en
+license:
+  - cc-by-4.0
+  - cc-by-sa-4.0
+tags:
+  - chemistry
+  - spectroscopy
+  - infrared
+  - nmr
+  - structure-elucidation
+  - cheminformatics
+size_categories:
+  - 100K<n<1M
+pretty_name: IRexp
+dataset_info:
+  features:
+    - name: id
+      dtype: string
+      description: InChIKey (full)
+    - name: inchikey
+      dtype: string
+    - name: smiles
+      dtype: string
+    - name: selfies
+      dtype: string
+    - name: ir_bands_cm-1
+      sequence: float64
+      description: Experimental IR wavenumbers (cm⁻¹), author-transcribed or peak-picked
+    - name: h_nmr
+      dtype: string
+      description: ¹H shift list as printed in the source paper
+    - name: c_nmr
+      dtype: string
+      description: ¹³C shift list as printed in the source paper
+    - name: ir_source
+      dtype: string
+      description: Always "experimental"
+    - name: source_doi
+      dtype: string
+      description: DOI or accession of the source publication
+    - name: license
+      dtype: string
+      description: CC-BY-4.0 (PMC) or CC-BY-SA-4.0 (Chemotion)
+---
+
+# IRexp — experimental IR band lists from open-access literature
+
+**Paper:** [IRSpectra-Bench and IRexp: candidate recall, not verification, limits LLM elucidation from real experimental IR and NMR](https://github.com/IlkhamFY/spectro-agent) (*Digital Discovery*, 2026)
+
+IRexp is the largest **openly redistributable** collection of **experimental infrared band lists** mined from open-access chemistry papers, often with co-reported ¹H/¹³C shift lists and resolved structures.
+
+> **Important:** IRexp contains **band lists** (peak positions in cm⁻¹), not digitised absorbance traces. This is the form reported in publication text — the regime IRSpectra-Bench evaluates — and is not directly comparable to SDBS or NIST full spectra.
+
+## Dataset summary
+
+| Split / file | Records | Description |
+|---|---:|---|
+| `irexp.jsonl.gz` | 121,233 | All IR band-list records |
+| `irexp_resolved.jsonl.gz` | 43,060 | Structure-linked (100%) |
+| … full IR + ¹H + ¹³C + structure | 33,201 | Multimodal quadruples |
+| `train_no_bench.jsonl.gz` | 42,808 | **Recommended for training** — `irexp_resolved` minus all IRSpectra-Bench InChIKey-14 |
+| `train_no_bench_nmr.jsonl.gz` | 32,949 | Same, requiring both ¹H and ¹³C |
+
+**Provenance:** 119,345 records from PMC Open-Access (CC-BY-4.0); 1,888 from Chemotion/RADAR4Chem (CC-BY-SA-4.0). Use `scripts/split_license_pools.py` to separate pools.
+
+**Companion benchmark:** [IRSpectra-Bench](https://github.com/IlkhamFY/spectro-agent/blob/main/docs/LEADERBOARD.md) — 194 blind elucidation problems built from IRexp; score submissions with `scripts/score_submission.py`.
+
+## Load in three lines
+
+```python
+from datasets import load_dataset
+
+# Full structure-linked corpus (43,060 records)
+ds = load_dataset("ilkhamfy/IRexp", data_files="data/irexp_resolved.jsonl.gz", split="train")
+
+row = ds[0]
+print(row["ir_bands_cm-1"][:5], row["smiles"][:40])
+```
+
+For **fine-tuning without benchmark leakage**, use `train_no_bench.jsonl.gz` (not the legacy `irexp_release/train.jsonl.gz`, which overlaps 117/200 benchmark compounds):
+
+```python
+ds = load_dataset("ilkhamfy/IRexp", data_files="data/train_no_bench.jsonl.gz", split="train")
+```
+
+Or clone the GitHub repo / Zenodo deposit and read locally:
+
+```python
+import gzip, json
+
+with gzip.open("irexp_resolved.jsonl.gz", "rt") as f:
+    row = json.loads(f.readline())
+```
+
+## Record schema
+
+Each JSONL row:
+
+```json
+{
+  "id": "AJCQUIFRMABSOZ-UHFFFAOYSA-N",
+  "inchikey": "AJCQUIFRMABSOZ-UHFFFAOYSA-N",
+  "smiles": "Cc1ccccc1NC(=O)Cn1cc...",
+  "selfies": "[C][C][=C]...",
+  "ir_bands_cm-1": [3318.0, 3146.0, 1704.0],
+  "h_nmr": "9.79 (s, 1H, NH-amide), ...",
+  "c_nmr": "164.87, 161.57, ...",
+  "ir_source": "experimental",
+  "source_doi": "10.1038/..."
+}
+```
+
+## Training vs benchmarking
+
+| Use case | File | Benchmark overlap |
+|---|---|---|
+| Pretrain IR encoder | `pretrain_ir.jsonl.gz` or all `ir_bands_cm-1` | N/A (mostly unlabeled) |
+| Supervised IR→structure | `train_no_bench.jsonl.gz` | **None** (248 IK-14 held out) |
+| Evaluate elucidation | [IRSpectra-Bench](https://github.com/IlkhamFY/spectro-agent/blob/main/docs/LEADERBOARD.md) | — |
+| ⚠️ Legacy split | `irexp_release/train.jsonl.gz` | **117/200 IK-14 overlap** — do not use for benchmark evaluation |
+
+Rebuild the held-out training pool:
+
+```bash
+python scripts/build_train_no_bench.py              # 42,808 rows
+python scripts/build_train_no_bench.py --require-nmr  # 32,949 rows (H+C required)
+```
+
+## Limitations (read before citing)
+
+- **Band lists, not spectra** — median 9 bands (PMC) vs 39 (Chemotion peak-picked).
+- **Literature-transcribed** — heterogeneous labs/instruments; not raw `.jdx` files.
+- **Structure resolution 35%** of all records; use `irexp_resolved` for supervised tasks.
+- **Extraction recall** of IR strings per paper not yet human-audited (transcription fidelity audited: 560/560 bands on n=60).
+
+## Citation
+
+```bibtex
+@article{yabbarov2026irspectra,
+  title   = {IRSpectra-Bench and {IRexp}: candidate recall, not verification,
+             limits {LLM} elucidation from real experimental {IR} and {NMR}},
+  author  = {Yabbarov, Ilkham and Sondhi, Rudra and Vargas-Hern{\'a}ndez, Rodrigo A.},
+  journal = {Digital Discovery},
+  year    = {2026},
+  note    = {Zenodo DOI to be assigned at publication}
+}
+```
+
+## Links
+
+- **Dataset (Hugging Face):** https://huggingface.co/datasets/ilkhamfy/IRexp
+- **Code & benchmark:** https://github.com/IlkhamFY/spectro-agent
+- **Leaderboard:** https://github.com/IlkhamFY/spectro-agent/blob/main/docs/LEADERBOARD.md
+- **Zenodo:** DOI minted at publication
+- **Licence details:** `NOTICE` in this repository (and `data/NOTICE` in the GitHub mirror)
+
+When uploading to Hugging Face, this file is the repository `README.md`.