---
license: cc-by-4.0
language:
  - en
tags:
  - routing
  - evaluation
  - llm-router
  - multi-domain
  - model-comparison
size_categories:
  - 1K<n<10K
configs:
  - config_name: results
    default: true
    data_files:
      - split: train
        path: data/results/train.jsonl.gz
  - config_name: verbose
    data_files:
      - split: train
        path: data/verbose/train.jsonl.gz
  - config_name: evals
    data_files:
      - split: train
        path: data/evals/train.jsonl.gz
---

# Dataset A - Routing (3 modelli, verdict-level)

**Totale query:** 5504 | **Gated (masked) query:** 0

Dataset di valutazione per LLM routing systems su 6 capability. Ogni query è stata eseguita su 3 modelli (qwen3.5-9b, deepseek-v4-flash, kimi2.6) e giudicata con grader deterministici (math/coding/ifeval) o LLM judge panel 2-of-3 (planning_agentic) / single judge (creative_synthesis, world_knowledge).

## Configs

```python
from datasets import load_dataset

# Pivot verdict per query (default config, 3 colonne bool)
ds_results = load_dataset("massaindustries/dataset-A-routing", split="train")

# Full: raw responses, costs, latency, 3x3 judge per planning, grader_meta
ds_verbose = load_dataset("massaindustries/dataset-A-routing", name="verbose", split="train")

# Prompts + ground truth (no model outputs) — per replicare i test
ds_evals = load_dataset("massaindustries/dataset-A-routing", name="evals", split="train")

# Tutti i config hanno una riga `_schema_anchor` (query_id == '_schema_anchor') con valori
# dummy per fissare lo schema di datasets/PyArrow. Filtrala via:
#   ds = ds.filter(lambda r: r["query_id"] != "_schema_anchor")
```

## Schema `results`

| campo | tipo | note |
|---|---|---|
| `query_id` | string | `q_NNNNN` |
| `query` | string | `<masked>` se gated |
| `dimension` | string | 1 di 6 capability |
| `evaluation_protocol_id` | string | protocollo grader |
| `source` | string | dataset originale |
| `gated` | bool | dataset proprietary (query mascherata) |
| `qwen_correct` | bool/null | verdict primario qwen3.5-9b |
| `ds4_correct` | bool/null | verdict primario deepseek-v4-flash |
| `kimi_correct` | bool/null | verdict primario kimi2.6 |

`*_correct` è `null` quando il judge ha astensione (~1.4% del totale).

## Schema `verbose`

Tutti i campi di `results` + per ogni modello m in {qwen, ds4, kimi}:
- `{m}_response` (string, raw output)
- `{m}_thinking` (string, raw thinking chain)
- `{m}_cost_usd` (float, null per qwen)
- `{m}_latency_ms` (int)
- `{m}_completion_tokens` (int)
- `{m}_reasoning_tokens` (int)
- `{m}_input_tokens` (int)
- `{m}_finish_reason` (string)
- `{m}_grader_meta` (string, JSON serializzato)
- `{m}_model_id_real` (string)
- `{m}_judge_gpt54mini` (bool/null, solo planning ST)
- `{m}_judge_mistral` (bool/null, solo planning ST)
- `{m}_judge_glm` (bool/null, solo planning ST)

## Verdict source per dimension

| dimension | rows | verdict source |
|---|---|---|
| planning_agentic | 1000 | panel 2-of-3 (gpt-5.4-mini + mistral-small-2603 + glm-5-turbo) per ST; single judge per MT |
| math_reasoning | 1000 | deterministic (math_equiv, gsm8k_final_answer) |
| coding | 1000 | deterministic (unit_test_pass) |
| instruction_following | 841 | deterministic (ifeval_constraint_check) |
| world_knowledge | 802 | mix (deterministic mcq_letter per 102; LLM judge `llm_judge_factual` per 700) |
| creative_synthesis | 696 | LLM judge `rubric_judge` (gpt-5.4-mini) |

## Win-rate per modello

| modello | correct | incorrect | abstention | accuracy |
|---|---|---|---|---|
| qwen | 3478 | 1786 | 241 | 0.6607 |
| ds4 | 4057 | 1220 | 228 | 0.7688 |
| kimi | 4130 | 784 | 591 | 0.8405 |

## Coverage check

Ogni modello ha verdict per tutte le 5339 query:

| modello | query con verdict | coverage |
|---|---|---|
| qwen | 5339 | 97.0% |
| ds4 | 5339 | 97.0% |
| kimi | 5339 | 97.0% |

## Reproducibility

- Stesso pool di 5339 query del dataset base [`massaindustries/dataset-A-routing-eval`](https://huggingface.co/datasets/massaindustries/dataset-A-routing-eval)
- Inference deterministic (temperature=0.0)
- Judge panel costanti, prompt template versionati nello skill `llmevals`
- Per ripetere i test: scarica config `evals`, esegui inference, applica grader

## License

CC-BY-4.0 (eccetto query originali dei dataset gated, che restano sotto le license dei rispettivi source).

## Citation

Per i source originali del dataset_A, vedere `lockfile.yaml` in [`massaindustries/dataset-A-routing-eval`](https://huggingface.co/datasets/massaindustries/dataset-A-routing-eval).