update dataset card

README.md

@@ -1,64 +1,125 @@
 ---
-license: apache-2.0
-dataset_info:
-  features:
-  - name: task_id
-    dtype: string
-  - name: paper_id
-    dtype: string
-  - name: analysis_target
-    dtype: string
-  - name: signal_model
-    dtype: string
-  - name: observable
-    dtype: string
-  - name: observable_pretty
-    dtype: string
-  - name: plot_units
-    dtype: string
-  - name: score_mode
-    dtype: string
-  - name: tolerance
-    dtype: float64
-  - name: walltime
-    dtype: string
-  - name: difficulty
-    dtype: string
-  - name: tags
-    list: string
-  - name: instructions_md
-    dtype: string
-  - name: task_toml
-    dtype: string
-  - name: template_yaml
-    dtype: string
-  - name: n_bins
-    dtype: int64
-  - name: paper_pdf
-    dtype: binary
-  - name: paper_pdf_sha256
-    dtype: string
-  - name: paper_pdf_bytes
-    dtype: int64
-  - name: object_efficiencies
-    list:
-    - name: filename
-      dtype: string
-    - name: data
-      dtype: binary
-    - name: sha256
-      dtype: string
-    - name: size_bytes
-      dtype: int64
-  splits:
-  - name: train
-    num_bytes: 7635049
-    num_examples: 10
-  download_size: 6722274
-  dataset_size: 7635049
-configs:
-- config_name: default
-  data_files:
-  - split: train
-    path: data/train-*
+license: mit
+task_categories:
+  - text-generation
+  - question-answering
+language:
+  - en
+tags:
+  - physics
+  - high-energy-physics
+  - particle-physics
+  - LHC
+  - CMS
+  - benchmark
+  - agentic
+  - llm-agents
+  - tool-use
+  - simulation
+pretty_name: Collider-Bench
+size_categories:
+  - n<1K
 ---
+
+# Collider-Bench
+
+**Collider-Bench** is a benchmark for evaluating whether LLM agents can reproduce experimental analyses from the Large Hadron Collider (LHC) using only public papers and open scientific software. Such analyses are often difficult to reproduce because the public toolchain only approximates the software used internally by the experimental collaborations, while the published papers inevitably omit implementation details needed for a faithful reconstruction. Agents must therefore rely on physical reasoning, domain knowledge, and trial-and-error to fill these gaps. Each task requires the agent to turn a published analysis into an executable simulation-and-selection pipeline and submit predicted collision event yields in specified signal regions.
+
+This HuggingFace dataset hosts the **task corpus only** — the agent-facing instructions, the null-filled HEPData-style template the agent fills, the CMS paper PDF, and the published object-efficiency maps. The **runtime harness, scorer, and hidden reference values** live in the companion GitHub repository:
+
+🔗 **https://github.com/dfaroughy/Collider-Bench**
+
+The reference yields used by the scorer are deliberately not published here — leaking them would let any LLM ingesting HF datasets memorize the answers and defeat the benchmark's blind-test property.
+
+## Quick start
+
+```python
+from datasets import load_dataset
+
+ds = load_dataset("Dariusfar/ColliderBench", split="train")
+print(ds)                                  # 10 sim tasks
+print(ds[0]["task_id"], ds[0]["paper_id"])
+print(ds[0]["instructions_md"][:400])      # what the agent gets shown
+print(ds[0]["template_yaml"][:400])        # the null-filled template
+print(ds[0]["paper_pdf"][:8])              # PDF magic bytes (b'%PDF-1.5')
+```
+
+To actually run an agent against a task and score its submission, install the harness from the GitHub repo:
+
+```bash
+git clone https://github.com/dfaroughy/Collider-Bench.git
+cd Collider-Bench
+pip install -e ".[dev]"
+podman pull ghcr.io/dfaroughy/lhc-bench:latest      # MadGraph + Pythia + Delphes + ROOT image
+export ANTHROPIC_API_KEY=...                         # or OPENAI_API_KEY / GEMINI_API_KEY / DEEPSEEK_API_KEY
+scripts/run-agent --config configs/anthropics/claude_sonnet.yaml --task sus-16-046_sim-T5Wg
+```
+
+## Schema (one row per task)
+
+| Field | Type | Description |
+|---|---|---|
+| `task_id`            | string | Canonical task identifier, e.g. `sus-16-046_sim-T5Wg` |
+| `paper_id`           | string | CMS analysis identifier, e.g. `CMS-SUS-16-046` |
+| `analysis_target`    | string | Final state under study (`photons`, `single lepton`, `leptons + jets`) |
+| `signal_model`       | string | SUSY simplified-model name + slice (`T5Wg, high-H_T`, `T2tt, compressed`, …) |
+| `observable`         | string | Observable key as used by the scorer (`STgamma`, `MET`, …) |
+| `observable_pretty`  | string | LaTeX-style label (`S_T^gamma`, `E_T^miss`, `p_T^miss`) |
+| `plot_units`         | string | y-axis units of the histogram (`Events/bin`, `Events/GeV`) |
+| `score_mode`         | string | Scoring mode used by the harness (`shape_norm` for sim tasks) |
+| `tolerance`          | float64 | Per-bin tolerance band used in shape-pass-rate diagnostics |
+| `walltime`           | string | Harness walltime budget (e.g. `2h30m`) |
+| `difficulty`         | string | `easy` / `medium` / `hard` rough difficulty tag |
+| `tags`               | seq<string> | Free-form metadata tags |
+| `instructions_md`    | string | The full `TASK.md` text — the agent's primary instructions |
+| `task_toml`          | string | Verbatim `task.toml` content (paper id, observable, walltime, tolerance, …) |
+| `template_yaml`      | string | Null-filled HEPData-style YAML the agent fills with predicted bin values |
+| `n_bins`             | int64  | Total bin count across all dependent variables in the template |
+| `paper_pdf`          | binary | Bytes of the CMS analysis paper (publicly available on CDS/Inspire-HEP) |
+| `paper_pdf_sha256`   | string | SHA-256 of `paper_pdf` |
+| `paper_pdf_bytes`    | int64  | Length of `paper_pdf` |
+| `object_efficiencies`| seq<{filename, data, sha256, size_bytes}> | CMS public detector-efficiency maps (ROOT files) the agent needs to apply during selection |
+
+## Task corpus
+
+| Task id | Analysis target | Signal | Observable | Paper |
+|---|---|---|---|---|
+| `sus-16-034_sim-TChiWZ`        | leptons + jets | `TChiWZ`              | $E_T^{\rm miss}$ | CMS-SUS-16-034 |
+| `sus-16-046_sim-T5Wg`          | photons        | `T5Wg`                | $S_T^{\gamma}$   | CMS-SUS-16-046 |
+| `sus-16-046_sim-TChiWg`        | photons        | `TChiWg`              | $S_T^{\gamma}$   | CMS-SUS-16-046 |
+| `sus-16-047_sim-T5Wg_highHT`   | photons        | `T5Wg`, high-$H_T$    | $p_T^{\rm miss}$ | CMS-SUS-16-047 |
+| `sus-16-047_sim-T5Wg_lowHT`    | photons        | `T5Wg`, low-$H_T$     | $p_T^{\rm miss}$ | CMS-SUS-16-047 |
+| `sus-16-047_sim-T6gg_highHT`   | photons        | `T6gg`, high-$H_T$    | $p_T^{\rm miss}$ | CMS-SUS-16-047 |
+| `sus-16-047_sim-T6gg_lowHT`    | photons        | `T6gg`, low-$H_T$     | $p_T^{\rm miss}$ | CMS-SUS-16-047 |
+| `sus-16-051_sim-T2tt_SRG`      | single lepton  | `T2tt`                | $E_T^{\rm miss}$ | CMS-SUS-16-051 |
+| `sus-16-051_sim-T2bW_SRG`      | single lepton  | `T2bW`                | $E_T^{\rm miss}$ | CMS-SUS-16-051 |
+| `sus-16-051_sim-T2tt_comp`     | single lepton  | `T2tt`, compressed    | $E_T^{\rm miss}$ | CMS-SUS-16-051 |
+
+## Scoring
+
+Each `sim` task asks the agent to reproduce the published per-bin yield distribution. The primary metric used by the scorer is the relative L² distance
+
+$$d(\hat y, y^\star) = \sqrt{\sum_k (\hat y_k - y_k^\star)^2 \big/ \sum_k (y_k^\star)^2}$$
+
+between the agent's bin yields $\hat y$ and the published reference $y^\star$, plus the integrated yield error $\Delta = |\Sigma\hat y - \Sigma y^\star| / \Sigma y^\star$. Diagnostic metrics (RMSLE, Jensen-Shannon, Baker-Cousins shape p-value) are also computed per run.
+
+Scoring is offline and deterministic — it does **not** require an LLM. See [`ColliderBench/Evals/`](https://github.com/dfaroughy/Collider-Bench/tree/main/ColliderBench/Evals) in the harness repo.
+
+## Citation
+
+If you use Collider-Bench in your research, please cite:
+
+```
+@misc{colliderbench2026,
+  title  = {Collider-Bench: A benchmark for LHC analysis recasting by LLM agents},
+  author = {Faroughy, Darius A. and contributors},
+  year   = {2026},
+  url    = {https://huggingface.co/datasets/Dariusfar/ColliderBench},
+}
+```
+
+…and the four underlying CMS papers (CMS-SUS-16-034, -046, -047, -051) as listed in the GitHub repo's [References section](https://github.com/dfaroughy/Collider-Bench#references).
+
+## License
+
+MIT (matches the GitHub repo). The CMS paper PDFs and detector efficiency maps are reproduced here as published by the CMS Collaboration under the terms of their respective public-data policies.