File size: 4,964 Bytes
694545c
 
 
 
0400752
694545c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0400752
 
694545c
fd69a65
0400752
694545c
 
 
0400752
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
# IRSpectra-Bench leaderboard

Blind structure elucidation from **molecular formula + IR + ¹H + ¹³C** peak lists exactly as reported in open-access papers. Constitution scoring uses RDKit InChIKey connectivity (first 14 characters) unless noted.

**Paper:** [IRexp and IRSpectra-Bench](https://github.com/IlkhamFY/spectro-agent) (manuscript in preparation, 2026).

---

## Main benchmark (n = 194)

| Rank | Model / method | Top-1 ↑ | Recall (top-3) ↑ | Gen. recall | Verif. prec. \| recall | Notes |
|---:|---|--:|--:|--:|--:|---|
| 1 | Claude Fable 5 | **46%** | 54% | — | — | 24-compound subset only |
| 2 | Claude Opus + generate-wide + forward-verify | **30%** | — | 42% | 72% | 60-compound arm |
| 3 | Claude Opus + forward-verify | 30% | 33.5% | 34% | **89%** | Full benchmark (headline) |
| 4 | Claude Opus (solver self-rank) | 28.4% | 33.5% | 34% | 85% | Full benchmark |
| 5 | Grok 4.6 | — | — | 53% | 62% | 60-compound arm |
| 6 | Gemini 3.7 Flash | — | — | 50% | 73% | 60-compound arm |
| 7 | GPT-5.6 Sol | — | — | 42% | 68% | 60-compound arm |
| 8 | Claude Sonnet | 21% | 25% | — | — | 24-compound subset |
| 9 | Claude Haiku | 0% | 4% | — | — | 24-compound subset |

Bootstrap 95% CIs for the headline row: top-1 **28.4% [22–35]**, recall **33.5% [27–40]**. Corpus-reweighted top-1 (17.5% simple / 82.5% complex): **15.2% [11–20]**.

**Key finding:** verification precision exceeds generation recall for every vendor tested — the binding constraint is *candidate proposal*, not spectral ranking.

### By difficulty (Claude Opus, n = 194)

| Stratum | n | Top-1 | Recall |
|---|---:|--:|--:|
| All | 194 | 28.4% | 33.5% |
| Simple | 98 | 48.0% | 54.1% |
| Complex | 96 | 8.3% | 12.5% |

---

## Evaluate your model

### 1. Download the benchmark (questions only — no answers in the solver prompt)

```bash
git clone https://github.com/IlkhamFY/spectro-agent.git
cd spectro-agent
pip install -r requirements.txt
```

Questions (blind inputs):

- `data/benchmark_main/questions2.jsonl` (140; use `clean_qids.json` for validated subset)
- `data/benchmark_v3/questions2.jsonl` (40)
- `data/benchmark_v2_ctrl/questions2.jsonl` (20)

Each row: `qid`, `formula`, `ir_bands_cm-1`, `h_nmr`, `c_nmr`. **No structure hints.**

### 2. Run your elucidator

Return up to **three ranked SMILES** per `qid`. Protocol:

- Inputs: formula + IR + ¹H + ¹³C only (as printed in the source paper).
- No web search, no structure hints, no answer-key access.
- Document model version, prompt, and tool access in your submission.

### 3. Score locally

Write predictions as JSONL:

```json
{"qid": "R01", "candidates": ["SMILES_rank1", "SMILES_rank2", "SMILES_rank3"]}
```

```bash
python scripts/score_submission.py --predictions my_run.jsonl --name "YourModel-1.0"
# optional strict stereochemistry scoring:
python scripts/score_submission.py --predictions my_run.jsonl --stereo
```

Reproduce the official headline numbers:

```bash
python scripts/score_main.py
python scripts/forward_verify_all.py
```

### 4. Submit to the leaderboard

Open a GitHub issue or PR on [IlkhamFY/spectro-agent](https://github.com/IlkhamFY/spectro-agent) with:

1. `--name` label for the table
2. `score_submission.py` output (copy-paste)
3. Predictions file (`my_run.jsonl`) or link to reproducible run
4. Model ID, date, and brief protocol note (tools, candidate budget, reasoning tier)
5. Confirmation: blind protocol, no answer-key access

We will verify scoring with `scripts/score_submission.py` before adding a row.

---

## Subsets & extensions

| Benchmark | n | Purpose |
|---|---:|---|
| **IRSpectra-Bench** (main + v3 + v2_ctrl) | 194 | Headline leaderboard |
| IRSpectra-Bench (main clean only) | 134 | Spectrally validated main round |
| IRSpectra-Bench-Electrolyte | 46 | Battery-electrolyte functional classes |
| Cross-vendor arm | 60 | Same compounds, multiple vendors (`docs/CROSS_VENDOR.md`) |
| Model comparison subset | 24 | Claude Haiku → Fable ladder |

---

## Related resources

- **IRexp dataset (training):** https://huggingface.co/datasets/ilkhamfy/IRexp — use `data/train_no_bench.jsonl.gz` to avoid benchmark leakage (`data/irexp_release/README_HF.md`)
- **Cross-vendor protocol:** `docs/CROSS_VENDOR.md`
- **Forward-verification:** `docs/FORWARD_VERIFY.md`
- **Full reproduction:** `README.md` in repository root

---

## Citation

If you use IRSpectra-Bench or report numbers on it, please cite:

```bibtex
@article{yabbarov2026irspectra,
  title   = {{IRexp} and {IRSpectra-Bench}: redistributable experimental {IR} band lists,
             a blind peak-list benchmark, and a recall-bound diagnosis of {LLM} elucidation},
  author  = {Yabbarov, Ilkham and Sondhi, Rudra and Vargas-Hern{\'a}ndez, Rodrigo A.},
  year    = {2026},
  note    = {Manuscript in preparation; target J. Chem. Inf. Model.}
}
```

*Last updated: 2026-08-25 (JCIM editorial strengthen). External submissions listed after verification.*