Datasets:

CosmosMind
/

ChemArgus

@@ -1,72 +1,52 @@
 ---
-task_categories:
-- question-answering
-language:
-- en
-tags:
-- chemistry
-- multimodal
-- competition-level
-- evaluation
-- benchmark
-pretty_name: ChemArgus
-size_categories:
-- n<1K
-configs:
-- config_name: dev
-  data_files:
-  - split: dev
-    path: dev/dataset.jsonl
-- config_name: test
-  data_files:
-  - split: test
-    path: test/dataset.jsonl
 ---
-# ChemArgus
-ChemArgus is a competition-level, multimodal, diagnostic benchmark for expert
-chemistry reasoning. Problems are drawn from chemistry competition material,
-textbooks, and research papers, and are scored by fine-grained rubrics with
-partial credit rather than by answer endpoints.
-Chemistry evaluation has moved past multiple-choice probes toward open-ended
-expert problems, yet the benchmarks that define the current frontier still score
-answer endpoints alone. Such scoring conflates memorized facts with stepwise
-derivation, treats visual and textual inputs as interchangeable, and confounds
-long-horizon planning with stepwise execution. ChemArgus replaces the answer
-endpoint with three measurement properties:
-- **Fine-grained scoring** grades every subquestion point by point with partial
-  credit, so a score lands at the level of the step rather than the outcome.
-- **Multimodal capability** is priced on every problem containing a real image,
-  paired with a controlled text description, so visual understanding is measured
-  against text on identical tasks.
-- **Diagnostic attribution** reads paired metrics that share one scoring unit and
-  differ only in context construction, so a gap between two scores isolates a
-  single cause: integration, propagation, or local reasoning.
-Reference scores, model results, and traces are intentionally excluded from this
-release.
-## Dataset partitions
-The 191 problems (23 exams) are divided into three partitions. All three
-partitions are drawn by seeded stratified sampling so that the distribution of
-**level** (exam series), **subfield** (major knowledge category), and
-**modality** (with / without model-visible images) is consistent across
-partitions and matches the full set.
-| Partition | Share | Problems | Subquestions | Marks | Content |
-| --- | ---: | ---: | ---: | ---: | --- |
-| `dev/` — Public Dev Set | 25% | 48 | 210 | 636 | problems **with** rubrics, for debugging |
-| `test/` — Held-out Test Set | 55% | 105 | 445 | 1387 | problems **without** rubrics; graded by the evaluation side |
-| `private/` — Private Set | 20% | 38 | 165 | 559 | never published (anti-cheating final check) |
-`private/` is not part of this repository. The evaluation side keeps it
-separately; `test/` submissions are scored against the held-out rubrics there.
-## Layout
 ```
 dev/
@@ -83,7 +63,7 @@ test/
     images/*.png
 ```
-## Data format
 Each problem is a JSON object:
@@ -112,48 +92,49 @@ Each problem is a JSON object:
 }
 ```
-- `problem_id`: `Exam-<nn>-T<task>` (`nn` is the 01-23 exam index);
-  `sub_id`: `{problem_id}-{n}` (sequential).
-- `header` is the shared background of the whole problem.
-- `text_open` is the subquestion text; `full_mark` is the integer maximum score.
-- In `dev/` problems, `rubric` lists the scoring points with partial-credit
-  annotations and `rubric_images` are the private scoring figures.
-  In `test/` problems the `rubric` and `rubric_images` fields are **absent**.
-- Image files use the unified naming scheme
-  `{problem_id}-{sub_index}-{n}-prompt|rubric.png` (subquestion images) and
-  `{problem_id}-header-{n}.png` (problem-level images).
-### Image visibility
-- `header_images` and `text_images` are **model-visible**.
-- `rubric_images` are **private scoring images**: they are only given to the
-  judge, never to the answering model.
-## Evaluation protocols
 Three protocols share one scoring unit and differ only in context construction:
-- **`full_problem`** — the model receives the complete problem (header, images,
-  all subquestions) in one call and answers everything in one response.
-- **`sequential_carry`** — the model answers subquestions in order; each answer
-  is carried into the next step, with no access to gold answers or rubrics.
-- **`oracle_scaffolded`** — the model answers each subquestion after being given
-  the correct result of the previous one; it isolates local reasoning from
-  error propagation.
-## Scoring
-The core metric is the **Normalized Rubric Score**: the rubric-grounded points
 awarded divided by the problem's total marks. Every subquestion is graded by an
 LLM judge against the private rubric (partial credit per scoring point), with
-structure/SMILES equivalence handled by chemical meaning. The judge prompt is
-specified in `harness/prompts/judge_rubric.md`.
-## Leaderboard
 Live results: <https://www.cosmosmind.ai/leaderboard/chemargus>
-## Citation
 ```bibtex
 @misc{chemargus2026,

+# 🔬 ChemArgus
+### From Answer Accuracy to Rubric-Grounded Reasoning Diagnosis<br>in Multimodal Expert-Level Chemistry
+![dev](https://img.shields.io/badge/dev-48%20with%20rubrics-brightgreen)
+![test](https://img.shields.io/badge/test-105%20held--out-orange)
+![private](https://img.shields.io/badge/private-38%20never%20published-lightgrey)
+![multimodal](https://img.shields.io/badge/input-multimodal%20%2B%20text-blueviolet)
+![language](https://img.shields.io/badge/language-en-blue)
 ---
+**ChemArgus** is a competition-level, multimodal, diagnostic benchmark for expert
+chemistry reasoning. Problems are drawn from chemistry competition material,
+textbooks, and research papers — and scored by fine-grained rubrics with partial
+credit, instead of answer endpoints.
+> Chemistry evaluation has moved past multiple-choice probes toward open-ended
+> expert problems, yet current frontier benchmarks still score answer endpoints
+> alone. Such scoring conflates memorized facts with stepwise derivation, treats
+> visual and textual inputs as interchangeable, and confounds long-horizon
+> planning with stepwise execution. ChemArgus replaces the answer endpoint with
+> three measurement properties:
+>
+> - **Fine-grained scoring** — every subquestion is graded point by point with
+>   partial credit, so a score lands at the level of the step.
+> - **Multimodal capability** — every problem with a real image is paired with a
+>   controlled text description, pricing visual understanding against text.
+> - **Diagnostic attribution** — paired metrics share one scoring unit and differ
+>   only in context construction, isolating integration, propagation, or local
+>   reasoning as the cause of a score gap.
 ---
+## 📊 Dataset Partitions
+| | Share | Problems | Subquestions | Marks | Public images | Rubrics |
+| --- | ---: | ---: | ---: | ---: | ---: | --- |
+| 🧪 **dev** · Public Dev Set | 25% | 48 | 210 | 636 | 268 | ✅ included |
+| 🏁 **test** · Held-out Test Set | 55% | 105 | 445 | 1387 | 268 | ❌ stripped |
+| 🔒 **private** · Private Set | 20% | 38 | 165 | 559 | — | 🔐 local only |
+| **Total** | 100% | **191** | **820** | **2582** | **536** | |
+> 🎲 **Seeded stratified split** — the distribution of *level* (exam series),
+> *subfield* (knowledge category), and *modality* (with / without model-visible
+> images) stays consistent across all three partitions and matches the full set
+> (max deviation ≈ 3 percentage points).
+## 📁 Repository Layout
 ```
 dev/
     images/*.png
 ```
+## 🧬 Data Format
 Each problem is a JSON object:
 }
 ```
+| Field | Meaning |
+| --- | --- |
+| `problem_id` | `Exam-<nn>-T<task>` — `nn` is the 01–23 exam index |
+| `sub_id` | `{problem_id}-{n}` (sequential within the problem) |
+| `header` | Shared background of the whole problem |
+| `text_open` | Subquestion text |
+| `full_mark` | Integer maximum score of the subquestion |
+| `rubric` | Scoring points with partial-credit annotations — **absent in `test/`** |
+| `*_images[].file` | Image file name (see naming scheme below) |
+| `*_images[].description` | Text description of the image content |
+Image files follow a unified naming scheme:
+`{problem_id}-{sub_index}-{n}-prompt|rubric.png` (subquestion images) and
+`{problem_id}-header-{n}.png` (problem-level images).
+> ⚠️ **Image visibility** — `header_images` and `text_images` are
+> **model-visible**; `rubric_images` are **private scoring images**, given only
+> to the judge, never to the answering model. They do not appear in `test/` at all.
+## 🔀 Evaluation Protocols
 Three protocols share one scoring unit and differ only in context construction:
+1. **`full_problem`** — the model receives the complete problem in one call and
+   answers everything in one response.
+2. **`sequential_carry`** — subquestions are answered in order; each answer is
+   carried into the next step, with no access to gold answers.
+3. **`oracle_scaffolded`** — each subquestion is answered after being given the
+   correct result of the previous one; isolates local reasoning from error
+   propagation.
+## ⚖️ Scoring
+The core metric is the **Normalized Rubric Score**: rubric-grounded points
 awarded divided by the problem's total marks. Every subquestion is graded by an
 LLM judge against the private rubric (partial credit per scoring point), with
+structure/SMILES equivalence judged by chemical meaning.
+## 🏆 Leaderboard
 Live results: <https://www.cosmosmind.ai/leaderboard/chemargus>
+## 📖 Citation
 ```bibtex
 @misc{chemargus2026,