| --- |
| license: apache-2.0 |
| task_categories: |
| - question-answering |
| - text-generation |
| language: |
| - en |
| tags: |
| - llm-routing |
| - reinforcement-learning |
| - agent |
| size_categories: |
| - 10K<n<100K |
| extra_gated_prompt: >- |
| These files contain questions converted from existing benchmarks (MATH, |
| Omni-MATH, APPS, LiveCodeBench, TACO, SuperGPQA, SATBench, TriviaQA, PopQA, |
| SimpleQA-Verified, and AIME / GPQA-Diamond / BBH for the held-out sets). Each |
| source dataset keeps its own licence and terms, and several are evaluation |
| benchmarks whose value depends on not being absorbed into training corpora. |
| Access is gated for that reason, not to restrict research use. |
| extra_gated_fields: |
| Name: text |
| Affiliation: text |
| I will use these files for research and evaluation only: checkbox |
| I will follow each source benchmark's own licence and terms: checkbox |
| I will not redistribute these files or include them in model training data: checkbox |
| --- |
| # RouteWeaver — training and evaluation sets |
|
|
| The prepared datasets for **RouteWeaver: Weaving Mode Selection and Execution |
| into Unified LLM Routing**. They are the exact files the reported runs read, so |
| training and evaluation reproduce against the same questions in the same order. |
|
|
| - 📄 Paper: coming soon |
| - 💻 Code: https://github.com/LaughKing/RouteWeaver |
| - 🌐 Project page: https://laughking.github.io/RouteWeaver/ |
| - 🤖 Model: https://huggingface.co/e2rea1/RouteWeaver-4B |
| - 💸 Cost-aware variants: https://huggingface.co/e2rea1/RouteWeaver-4B-cost |
| - 🗂️ Everything on the Hub: https://huggingface.co/collections/e2rea1/routeweaver-6ac89596851c1c7650f72aa4 |
|
|
| ```bash |
| python data/hf_download.py # from the code repository |
| ``` |
|
|
| ## Contents |
|
|
| | File | Rows | What it is | |
| |---|---|---| |
| | `parquet/routeweaver_train.parquet` | 4,000 | the training recipe. Rows 0-799 are the cold start's own batches, verbatim; rows 800+ are the 3,200-question recipe (batch_index 25-124) | |
| | `parquet/routeweaver_cold_pool.parquet` | 4,672 | the pool the cold start trains on | |
| | `parquet/val_stub.parquet` | 32 | the validation stub verl requires even with validation off | |
| | `eval/eval_600_free.parquet` | 600 | MATH L2-5 and APPS come from here | |
| | `eval/eval_700_free.parquet` | 700 | Omni-MATH, LiveCodeBench, TACO, SuperGPQA, SATBench, SimpleQA-Verified | |
| | `eval/eval_triviaqa120_free.parquet` | 120 | TriviaQA | |
| | `eval/ood_aime149_free.parquet` | 149 | held out (Math) | |
| | `eval/ood_gpqa_diamond198_free.parquet` | 198 | held out (Knowledge) | |
| | `eval/ood_bbh1080_free.parquet` | 1,080 | held out (Reason) | |
| | `cost_calib/cost_calib_128.parquet` | 128 | the queries C_ref is calibrated on | |
| |
| The nine reported benchmarks (1,060 questions) are drawn from the three |
| in-distribution files: MATH and APPS from `eval_600`, TriviaQA from its own |
| file, the other six from `eval_700`. |
|
|
| `checksums.json` records each file's row count, per-dataset composition and |
| sha256. |
|
|
| ## Row format |
|
|
| Every row is one question in verl's prompt format: |
|
|
| - `prompt` — the chat message the router sees. The mode-selection menu and the |
| six anonymous worker ids with their capability descriptions are **baked in**, |
| so the prompt is fixed at build time and cannot drift with the code. |
| - `reward_model` — the gold answer or test cases the scorer judges against. |
| - `extra_info` — `sample_id`, `domain`, `dataset`, `batch_index` (the training |
| step the row belongs to; the loader reads the file with `shuffle=False`, so |
| row order IS the curriculum), `forced_mode`, and `payload_json` with the raw |
| question and its metadata. |
|
|
| `batch_index` drives the forced-to-free schedule: ρ_b falls from 1 at batch 25 |
| to 0 at batch 75, so a row's position in the file decides whether its mode is |
| assigned or chosen. |
| |
| ## Sources and licences |
| |
| Built from MATH, Omni-MATH, APPS, LiveCodeBench, TACO, SuperGPQA, SATBench, |
| TriviaQA, PopQA and SimpleQA-Verified, plus AIME, GPQA-Diamond and BBH for the |
| held-out sets. Each original dataset keeps its own licence and terms; this |
| release redistributes selected questions in a converted form for reproducing |
| the paper. Test questions do not overlap the training questions. |
| |
| ## Citation |
| |
| ```bibtex |
| @article{routeweaver2026, |
| title = {RouteWeaver: Weaving Mode Selection and Execution into Unified LLM Routing}, |
| author = {Wang, Xiaohan and Zhang, Haozhen and Liu, Qingyuan and Feng, Tao and Wang, Wenya}, |
| journal = {arXiv preprint}, |
| year = {2026} |
| } |
| ``` |
| |
