Dataset card v1.0

README.md

@@ -1,50 +1,98 @@
 ---
-dataset_info:
-  features:
-  - name: id
-    dtype: string
-  - name: source
-    dtype: string
-  - name: domain
-    dtype: string
-  - name: kind
-    dtype: string
-  - name: gold
-    dtype: string
-  - name: prompt
-    dtype: string
-  - name: has_image
-    dtype: bool
-  - name: est_tokens
-    dtype: int64
-  - name: prompt_truncated
-    dtype: bool
-  - name: router_input
-    dtype: string
-  - name: route
-    dtype: string
-  - name: agent_results
-    dtype: string
-  - name: messages
-    list:
-    - name: role
-      dtype: string
-    - name: content
-      dtype: string
-  splits:
-  - name: train
-    num_bytes: 19208201
-    num_examples: 4720
-  - name: test
-    num_bytes: 1974064
-    num_examples: 523
-  download_size: 8515194
-  dataset_size: 21182265
-configs:
-- config_name: default
-  data_files:
-  - split: train
-    path: data/train-*
-  - split: test
-    path: data/test-*
+license: cc-by-sa-4.0
+language: [en]
+task_categories: [text-classification]
+tags: [routing, orchestration, llm-router, multi-agent, model-selection]
+pretty_name: Shaktimaan Routing Data v1.0
+size_categories: [1K<n<10K]
 ---
+# Shaktimaan Routing Data v1.0
+
+The training and evaluation data behind [Shaktimaan v1.0](https://huggingface.co/podstack/Shaktimaan), a 0.5B
+LLM router. Each row is one task that **every capable agent in a six-model pool attempted**, with every answer
+automatically graded, plus the cost-aware routing label derived from those outcomes. Useful for training or
+benchmarking LLM routers, model-selection policies and cascades.
+
+Published: 4720 train / 523 test rows (5243
+tasks). The model was additionally trained and evaluated on 300 ChartQA rows that are **not** published here (see
+Licensing).
+
+## Agent pool
+| agent | model | context | image input | relative cost / token |
+|---|---|---|---|---|
+| `gemma-3-1b` | `google/gemma-3-1b-it` | 32k | no | 0.02 |
+| `kimi-vl` | `Kimi-VL-A3B-Thinking` | 8k | yes | 0.05 |
+| `nemotron-lightning` | `nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4` | 128k | no | 0.1 |
+| `glm-5.3-flash` | `zai-org/GLM-5.3-Flash` | 32k | no | 0.2 |
+| `qwen3.8-27b` | `Qwen/Qwen3.8-27B` | 32k | no | 0.5 |
+| `nemotron-ultra` | `nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4` | 256k | no | 1.0 |
+
+"Relative cost" is a per-token price weight used to compute routing cost; it reflects each model's relative serving
+cost (roughly active parameters), not a vendor price.
+
+## Sources
+| source | category | train | test | grading | license |
+|---|---|---|---|---|---|
+| [GSM8K](https://huggingface.co/datasets/openai/gsm8k) | math word problems | 895 | 105 | numeric match | MIT |
+| [MATH](https://huggingface.co/datasets/EleutherAI/hendrycks_math) | competition math | 894 | 106 | normalised LaTeX / numeric match | MIT |
+| [MMLU-Pro](https://huggingface.co/datasets/TIGER-Lab/MMLU-Pro) | knowledge MCQ (10 options) | 1080 | 120 | option letter | MIT |
+| [ARC-Challenge](https://huggingface.co/datasets/allenai/ai2_arc) | science MCQ | 725 | 75 | option letter | CC-BY-SA-4.0 |
+| [MBPP](https://huggingface.co/datasets/google-research-datasets/mbpp) | Python coding | 633 | 67 | unit tests executed in a network-less sandbox | CC-BY-4.0 |
+| [HumanEval](https://huggingface.co/datasets/openai/openai_humaneval) | Python coding | 146 | 18 | unit tests executed in a network-less sandbox | MIT |
+| [LongBench-v2](https://huggingface.co/datasets/THUDM/LongBench-v2) | long-context MCQ (up to ~260k tokens) | 347 | 32 | option letter | Apache-2.0 |
+| [ChartQA](https://huggingface.co/datasets/HuggingFaceM4/ChartQA) | chart images | 269 | 31 | relaxed match (numbers within 5%) | GPL-3.0 |
+
+Long LongBench-v2 contexts are stored truncated to what the router reads (first 5,000 and last 1,500 characters);
+`est_tokens` keeps the original size.
+
+## How it was built
+1. **Solve** - every agent answered every task once (temperature 0.6, up to 8,192 output tokens). Agents that cannot
+   take an input (no image support, context window too small) are recorded as failures without being called
+   (`skipped`: `no_vision` / `context`).
+2. **Grade** - automatically, per source (see table). Code is graded by running the dataset's unit tests against the
+   model's code in a throwaway container with no network, 512 MB RAM and a 15 s limit.
+3. **Label** - the cheapest agent that solved the task, counting a win only if that agent is reliable on the source
+   (solve rate >= 70%), because single-sample wins by weak agents are mostly luck a router cannot predict.
+   `gemma-3-1b` is never a label. If no eligible agent solved the task, the label is an ensemble of the capable strong
+   agents (`glm-5.3-flash`, `qwen3.8-27b`, `nemotron-ultra`, aggregated by the last), or the strongest capable agent
+   alone when fewer than two can take the input.
+
+### Agent solve rates (single attempt, all 5543 tasks incl. ChartQA)
+| source | `gemma-3-1b` | `kimi-vl` | `nemotron-lightning` | `glm-5.3-flash` | `qwen3.8-27b` | `nemotron-ultra` |
+|---|---|---|---|---|---|---|
+| GSM8K | 26.3% | 78.0% | 95.0% | 96.9% | 96.8% | 95.6% |
+| MATH | 18.4% | 41.9% | 70.7% | 75.8% | 75.9% | 71.6% |
+| MMLU-Pro | 12.7% | 50.1% | 73.0% | 81.0% | 82.5% | 76.2% |
+| ARC-Challenge | 44.9% | 91.4% | 97.9% | 97.8% | 98.0% | 98.2% |
+| MBPP | 50.4% | 67.0% | 90.9% | 94.0% | 88.4% | 93.6% |
+| HumanEval | 41.5% | 79.3% | 92.7% | 92.1% | 93.3% | 97.0% |
+| LongBench-v2 | 5.3% | 0.0% | 25.9% | 10.3% | 12.1% | 54.6% |
+| ChartQA | 0.0% | 80.0% | 0.0% | 0.0% | 0.0% | 0.0% |
+| **all** | **25.2%** | **60.8%** | **75.8%** | **78.1%** | **77.9%** | **79.3%** |
+
+### Label distribution
+| route label | tasks | share |
+|---|---|---|
+| `glm-5.3-flash` | 1854 | 33.4% |
+| `nemotron-lightning` | 1509 | 27.2% |
+| `kimi-vl` | 1373 | 24.8% |
+| `glm-5.3-flash,qwen3.8-27b,nemotron-ultra (ensemble)` | 365 | 6.6% |
+| `nemotron-ultra` | 329 | 5.9% |
+| `qwen3.8-27b` | 113 | 2.0% |
+
+## Fields
+| field | description |
+|---|---|
+| `id`, `source`, `domain`, `kind` | task id, source dataset, domain, grading kind (`num`, `math`, `mc`, `code`, `relaxed`) |
+| `prompt` | the task (LongBench contexts truncated, see above) |
+| `gold` | reference answer; for code, the unit tests |
+| `has_image`, `est_tokens`, `prompt_truncated` | input metadata (`est_tokens` is a conservative chars/3 estimate) |
+| `router_input` | exactly what the router reads: a size / image header plus the start and end of the task |
+| `agent_results` | JSON: per agent `correct`, `completion_tokens`, `latency` (s), `skipped` |
+| `route` | JSON label: `domain`, `difficulty`, `strategy` (`single` / `ensemble`), `agents` |
+| `messages` | chat-format SFT example (system, user = `router_input`, assistant = `route`) |
+
+## Licensing
+Released under **CC-BY-SA-4.0**, as required by the ARC-Challenge share-alike terms; the other included sources are
+MIT, CC-BY-4.0 and Apache-2.0. The 300 ChartQA rows used for training are excluded because ChartQA is GPL-3.0;
+`gen.py` in the model repository regenerates them. Please also cite the original datasets.