README: add viewer:false metadata (test disabling Dataset Viewer)
#4
by YunQ572 - opened
README.md
CHANGED
|
@@ -1,3 +1,7 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
# 🔬 ChemArgus
|
| 2 |
|
| 3 |
### From Answer Accuracy to Rubric-Grounded Reasoning Diagnosis<br>in Multimodal Expert-Level Chemistry
|
|
@@ -130,9 +134,49 @@ awarded divided by the problem's total marks. Every subquestion is graded by an
|
|
| 130 |
LLM judge against the private rubric (partial credit per scoring point), with
|
| 131 |
structure/SMILES equivalence judged by chemical meaning.
|
| 132 |
|
| 133 |
-
##
|
| 134 |
-
|
| 135 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 136 |
|
| 137 |
## 📖 Citation
|
| 138 |
|
|
|
|
| 1 |
+
---
|
| 2 |
+
viewer: false
|
| 3 |
+
---
|
| 4 |
+
|
| 5 |
# 🔬 ChemArgus
|
| 6 |
|
| 7 |
### From Answer Accuracy to Rubric-Grounded Reasoning Diagnosis<br>in Multimodal Expert-Level Chemistry
|
|
|
|
| 134 |
LLM judge against the private rubric (partial credit per scoring point), with
|
| 135 |
structure/SMILES equivalence judged by chemical meaning.
|
| 136 |
|
| 137 |
+
## 🏅 Results & Leaderboard
|
| 138 |
+
|
| 139 |
+
*Snapshot: September 2026 · Normalized Rubric Score (%). Runs are sorted by
|
| 140 |
+
`full_problem` within each input mode.*
|
| 141 |
+
|
| 142 |
+
🥇 **`full_problem` leaders** — Multimodal: Claude Opus 5 53.8%† ·
|
| 143 |
+
Text: Claude Opus 5 61.5%† · Text + reader: GLM-5.2 52.0%
|
| 144 |
+
|
| 145 |
+
| Model | Input | full_problem | sequential_carry | oracle_scaffolded |
|
| 146 |
+
| --- | --- | ---: | ---: | ---: |
|
| 147 |
+
| Claude Opus 5 † | multimodal | 53.8 | 59.4 | 64.9 |
|
| 148 |
+
| Gemini 3.8 Flash § | multimodal | 48.5 | 58.5 | 60.3 |
|
| 149 |
+
| Grok 4.6 § | multimodal | 33.8 | 41.8 | 45.5 |
|
| 150 |
+
| Qwen3.8-Max | multimodal | 29.3 | 31.2 | 34.0 |
|
| 151 |
+
| Kimi K3 | multimodal | 27.5 | 30.6 | 31.8 |
|
| 152 |
+
| MiniMax M3 | multimodal | 27.3 | 27.5 | 27.9 |
|
| 153 |
+
| GPT-5.6-Sol | multimodal | 23.2 | 28.8 | 30.4 |
|
| 154 |
+
| Claude Opus 5 † | text | 61.5 | 61.7 | 61.7 |
|
| 155 |
+
| Gemini 3.8 Flash § | text | 46.5 | 59.0 | 59.7 |
|
| 156 |
+
| Qwen3.8-Max | text | 36.0 | 39.0 | 41.9 |
|
| 157 |
+
| Kimi K3 | text | 34.5 | 36.0 | 37.5 |
|
| 158 |
+
| Grok 4.6 § | text | 33.4 | 42.0 | 46.0 |
|
| 159 |
+
| GPT-5.6-Sol | text | 29.8 | 35.6 | 36.5 |
|
| 160 |
+
| MiniMax M3 | text | 26.2 | 28.7 | 29.1 |
|
| 161 |
+
| DeepSeek V4 Pro | text | 25.0 | 25.0 | 27.6 |
|
| 162 |
+
| GLM-5.2 | text | 23.5 | 24.9 | 25.1 |
|
| 163 |
+
| GLM-5.2 | text + reader | 52.0 | 52.2 | 52.4 |
|
| 164 |
+
| Claude Opus 5 † | text + reader | 46.2 | 53.9 | 54.7 |
|
| 165 |
+
| DeepSeek V4 Pro | text + reader | 36.7 | 36.8 | 37.0 |
|
| 166 |
+
| Gemini 3.8 Flash § | text + reader | 33.8 | 45.1 | 49.7 |
|
| 167 |
+
| Kimi K3 | text + reader | 32.3 | 32.3 | 32.4 |
|
| 168 |
+
| GPT-5.6-Sol | text + reader | 31.1 | 33.6 | 34.0 |
|
| 169 |
+
| Grok 4.6 § | text + reader | 28.6 | 29.0 | 31.1 |
|
| 170 |
+
| Qwen3.8-Max | text + reader | 27.9 | 29.2 | 31.5 |
|
| 171 |
+
| MiniMax M3 | text + reader | 27.2 | 27.3 | 27.3 |
|
| 172 |
+
|
| 173 |
+
> † Claude Opus 5 runs rest on a small problem prefix (39–103 points) and are
|
| 174 |
+
> provisional until the full set completes.
|
| 175 |
+
> § Gemini 3.8 Flash and Grok 4.6 runs sit on a 12-paper, 1136-point subset.
|
| 176 |
+
> Percentages are computed on each run's own scored problem set, so
|
| 177 |
+
> denominators differ across runs.
|
| 178 |
+
|
| 179 |
+
Live leaderboard: <https://www.cosmosmind.ai/leaderboard/chemargus>
|
| 180 |
|
| 181 |
## 📖 Citation
|
| 182 |
|