Complete dataset card: licence, NOTICE, schema, benchmark links, citation

NOTICE.md

@@ -0,0 +1,17 @@
+# Notice
+
+`rewards.jsonl` contains measurements produced by Benchgen and is released under CC-BY-4.0.
+It contains no question text and no raw model output.
+
+Task identifiers refer to rows of upstream public datasets. Those datasets keep their own licences:
+
+| Dataset | Licence |
+|---|---|
+| HuggingFaceH4/MATH-500 | MIT (inherited from MATH; the mirror lists no licence field) |
+| TIGER-Lab/MMLU-Pro | MIT |
+| cais/mmlu | MIT |
+| allenai/ai2_arc | CC-BY-SA-4.0 |
+| openai/gsm8k | MIT |
+| opencompass/AIME2025 | MIT |
+
+Licences were read from each dataset's Hugging Face metadata on 2026-10-15. Check the upstream page before redistributing.

README.md

@@ -1,7 +1,6 @@
 ---
-license: other
-license_name: mixed-upstream
-license_link: https://github.com/tahsinsoyak/benchgen-router-dataset
+license: cc-by-4.0
+license_link: https://huggingface.co/datasets/benchgen/router-pilot/blob/main/NOTICE.md
 task_categories:
   - question-answering
   - text-classification
@@ -19,15 +18,13 @@ configs:
     data_files: rewards.jsonl
 ---
 
-# BenchGen Router Pilot: per-question agent rewards
+# Benchgen Router Pilot: per-question agent rewards
 
-A **routing** dataset: for each question, how every agent in a fixed pool actually performed.
-It is the input a model-selection policy trains on — not a question-answering dataset.
+A **routing** dataset: for each question, how every agent in a fixed pool actually performed. It is the input a model-selection policy trains on, not a question-answering dataset. Published by [Benchgen](https://benchgen.com/?utm_source=huggingface&utm_medium=dataset_card&utm_campaign=router-lite&utm_content=router-pilot-header), the benchmarking infrastructure for AI agents.
 
-Each row is one task. `mean_reward[i]` is the accuracy of agent `agent_order[i]` over
-3 independent attempts. Ties are preserved explicitly rather than collapsed by `argmax`,
-because on easy questions several agents are genuinely equal and pretending otherwise teaches a
-router a coin flip.
+Each row is one task. `mean_reward[i]` is the accuracy of agent `agent_order[i]` over 3 independent attempts. Ties are preserved explicitly rather than collapsed by `argmax`, because on easy questions several agents are genuinely equal and pretending otherwise teaches a router a coin flip.
+
+Companion dataset with the prompts: [`benchgen/router-pilot-tasks`](https://huggingface.co/datasets/benchgen/router-pilot-tasks). Join on `task_id`.
 
 ## Contents
 
@@ -37,7 +34,15 @@ router a coin flip.
 | --- | --- |
 | Source | HuggingFaceH4/MATH-500 14, TIGER-Lab/MMLU-Pro 10, allenai/ai2_arc 6, cais/mmlu 10, openai/gsm8k 4, opencompass/AIME2025 2 |
 | Domain | knowledge 20, math 20, reasoning 6 |
-| Difficulty | easy 9, hard 18, medium 19 |
+| Difficulty | easy 9, medium 19, hard 18 |
+
+## Schema (`rewards.jsonl`)
+
+| Field | Type | Notes |
+| --- | --- | --- |
+| `task_id` | string | Join key with `benchgen/router-pilot-tasks`, format `{source}/{split}/{index}` |
+| `agent_order` | list of string | The 5 agent ids, in the order used by every array below |
+| `mean_reward` | list of float | Accuracy of each agent over 3 attempts, 0.0 to 1.0 |
 
 ## The pool
 
@@ -49,10 +54,18 @@ router a coin flip.
 | 4 | `open_mid` | `mistralai/mistral-nemo` | direct | 0.03 |
 | 7 | `open_cheap_reasoning` | `inclusionai/ling-3.0-flash` | reasoning | 0.063 |
 
+## Provenance and licence
 
-## Collected but excluded
+This dataset contains **our own measurements** (graded correctness per agent per task). It contains no question text and no raw model output.
+
+| Part | Licence | Note |
+| --- | --- | --- |
+| Reward labels (`rewards.jsonl`) | CC-BY-4.0 | Measurements by Benchgen. Attribute "Benchgen" |
+| Task identity (`task_id`) | n/a | Identifies an upstream row, no upstream text is included |
 
-These were collected and are not in this release, because a label we cannot trust is worse than no label.
+Upstream sources the tasks come from (licence recorded 2026-10-15 from each dataset's Hugging Face metadata): `openbmb/RLPR-Train-Dataset` Apache-2.0 (excluded, see below), `HuggingFaceH4/MATH-500` MIT (inherited from MATH, the mirror lists no licence field), `TIGER-Lab/MMLU-Pro` MIT, `cais/mmlu` MIT, `allenai/ai2_arc` CC-BY-SA-4.0, `openai/gsm8k` MIT, `opencompass/AIME2025` MIT.
+
+## Collected but excluded
 
 | Source | Why |
 | --- | --- |
@@ -60,35 +73,48 @@ These were collected and are not in this release, because a label we cannot trus
 
 ## What is deliberately not here
 
-- **No question text.** Upstream licences differ per source: ARC-Challenge is share-alike, and
-  MMLU-Pro's licence was never verified. `task_id` identifies the upstream row so prompts can be
-  rehydrated from the original datasets.
-- **No raw completions.** Provider terms on republishing model outputs were not cleared. The
-  routing signal — correctness, ties, empties, errors — survives without them.
+- **No question text.** Upstream licences differ per source, and ARC-Challenge is share-alike. `task_id` identifies the upstream row so prompts can be rehydrated from the original datasets.
+- **No raw completions.** Provider terms on republishing model outputs were not cleared. The routing signal (correctness, ties, empties, errors) survives without them.
 
 ## Limitations, stated plainly
 
-- **High tie rate (43/46 tasks).** Most agents solve most of these questions, so the
-  per-question best agent is often arbitrary. A router trained on this alone learns little; the
-  dataset is most useful as evidence that routing does not pay on saturated benchmarks.
-- **One agent dominates.** `frontier_a` or `frontier_b` is the best agent on almost every
-  dataset here, so the pool behaves as a ranking rather than a set of specialists. That is a
-  property of the pool, and no amount of extra data changes it.
-- **The pool was hand-picked, not selected by measurement.** Trinity's Appendix A.6 chooses
-  agents and datasets jointly from a full screening matrix; that screening was skipped for cost.
-- **No direct-vs-reasoning contrast.** The intended pair of one base model in both modes was
-  dropped: every candidate `-thinking` slug ignored `max_tokens` (24k-75k tokens against a 4,096
-  cap), which breaks protocol comparability.
+- **High tie rate (43 of 46 tasks).** Most agents solve most of these questions, so the per-question best agent is often arbitrary. A router trained on this alone learns little. The dataset is most useful as evidence that routing does not pay on saturated benchmarks.
+- **One agent dominates.** `frontier_a` or `frontier_b` is the best agent on almost every dataset here, so the pool behaves as a ranking rather than a set of specialists. That is a property of the pool, and no amount of extra data changes it.
+- **The pool was hand-picked, not selected by measurement.** Trinity's Appendix A.6 chooses agents and datasets jointly from a full screening matrix. That screening was skipped for cost.
+- **No direct-vs-reasoning contrast.** The intended pair of one base model in both modes was dropped: every candidate `-thinking` slug ignored `max_tokens` (24k to 75k tokens against a 4,096 cap), which breaks protocol comparability.
 - **No code domain.** Execution-based grading needs a sandbox that does not exist here.
 - **Small.** This is a pilot, sized to a strict budget.
 
 ## Protocol
 
-Identical for every agent, because `E(D, M)` is not comparable otherwise: `max_tokens` 4096,
-`temperature` 0.1, `top_p` 1.0, minimal reasoning effort, 3 repetitions. Any response
-exceeding the token cap aborts collection rather than being recorded, so no row mixes protocols.
+Identical for every agent, because `E(D, M)` is not comparable otherwise: `max_tokens` 4096, `temperature` 0.1, `top_p` 1.0, minimal reasoning effort, 3 repetitions. Any response exceeding the token cap aborts collection rather than being recorded, so no row mixes protocols.
 
 ## Grading
 
-Answers are extracted by tagged answer, boxed LaTeX, labelled answer, then a last-number
-fallback that strips markup first. An empty response is never scored correct.
+Answers are extracted by tagged answer, boxed LaTeX, labelled answer, then a last-number fallback that strips markup first. An empty response is never scored correct.
+
+## Run the underlying benchmarks on Benchgen
+
+These are the public benchmarks the tasks were drawn from. Each has a page on Benchgen with a leaderboard and evaluations:
+
+| Source | Benchgen page |
+| --- | --- |
+| MATH-500 | [benchgen.com/benchmarks/lightman-et-al-openai/math-500](https://benchgen.com/benchmarks/lightman-et-al-openai/math-500?utm_source=huggingface&utm_medium=dataset_card&utm_campaign=router-lite&utm_content=math-500) |
+| MMLU-Pro | [benchgen.com/benchmarks/tiger-ai-lab-university-of-waterloo/mmlu-pro](https://benchgen.com/benchmarks/tiger-ai-lab-university-of-waterloo/mmlu-pro?utm_source=huggingface&utm_medium=dataset_card&utm_campaign=router-lite&utm_content=mmlu-pro) |
+| MMLU | [benchgen.com/benchmarks/hendrycks-et-al/mmlu](https://benchgen.com/benchmarks/hendrycks-et-al/mmlu?utm_source=huggingface&utm_medium=dataset_card&utm_campaign=router-lite&utm_content=mmlu) |
+| GSM8K | [benchgen.com/benchmarks/cobbe-et-al-openai/gsm8k](https://benchgen.com/benchmarks/cobbe-et-al-openai/gsm8k?utm_source=huggingface&utm_medium=dataset_card&utm_campaign=router-lite&utm_content=gsm8k) |
+| ARC-Challenge | [benchgen.com/benchmarks/clark-et-al-allen-institute-for-ai/arc-c](https://benchgen.com/benchmarks/clark-et-al-allen-institute-for-ai/arc-c?utm_source=huggingface&utm_medium=dataset_card&utm_campaign=router-lite&utm_content=arc-c) |
+| AIME 2025 | [benchgen.com/benchmarks/mathematical-association-of-america/aime-2025](https://benchgen.com/benchmarks/mathematical-association-of-america/aime-2025?utm_source=huggingface&utm_medium=dataset_card&utm_campaign=router-lite&utm_content=aime-2025) |
+
+To evaluate a router built from this data, see the [Router Fidelity Benchmark](https://benchgen.com/benchmarks/platform/router-fidelity-benchmark?utm_source=huggingface&utm_medium=dataset_card&utm_campaign=router-lite&utm_content=router-fidelity).
+
+## Citation
+
+```bibtex
+@misc{benchgen_router_pilot_2026,
+  title        = {Benchgen Router Pilot: per-question agent rewards},
+  author       = {{Benchgen}},
+  year         = {2026},
+  howpublished = {\url{https://huggingface.co/datasets/benchgen/router-pilot}}
+}
+```