sijirama/kairos-routing

Kairos Routing

Kairos Routing is a long-format dataset for training a model to select the best AI model for a prompt. Each row describes one candidate model evaluated on one prompt.

The intended learning problem is:

prompt + candidate_model_vector -> score

The router can score several candidate models for the same prompt and select the model with the highest predicted score.

Dataset Summary

• Approximately 1.92 million rows
• Approximately 144,000 unique prompts
• 64 candidate model routes
• 131 source tasks
• One row per prompt and candidate model pair
• Target scores are between 0.0 and 1.0 in the released data
• Stored as Parquet

Source Datasets

Kairos Routing combines three public datasets:

1. RouterBench

   RouterBench provides prompt-level evaluations across multiple candidate models. Its 0-shot and 5-shot configurations are converted from wide tables into one row per prompt and candidate model. The source model score becomes target_score. Observed model cost is retained when available.

2. SPROUT

   SPROUT provides prompts and per-model response records. Each response record includes a judge score and, where available, token counts and text. The judge score becomes target_score.

3. LLMRouterBench

   LLMRouterBench provides long-format records containing prompts, model outputs, scores, token counts, and observed costs. Its record score becomes target_score.

Row Structure

Each row contains source and audit fields such as:

• source
• source_config
• source_split
• source_task
• prompt_id
• prompt
• golden_answer
• candidate_model
• target_score
• raw_score
• score_type
• prompt_tokens
• completion_tokens
• observed_cost

The joined registry also contains route metadata such as base_model_id, provider, and model_family. These fields are retained for auditing and analysis, but the Kairos model input does not use model identity or provider identity.

Model Vector

The training model uses the candidate model's capabilities rather than its name. The current vector contains:

• Input price per 1M tokens
• Output price per 1M tokens
• Context window
• Maximum output tokens
• Model size in billions of parameters
• Text support
• Vision support
• Audio support
• Tool-calling support
• P50 latency
• P95 latency
• Throughput in tokens per second
• MMLU-Pro score
• GPQA score
• HumanEval score
• GSM8K score
• IFEval score
• Arena Elo

The vector also contains missing-value indicators for fields that may have been unavailable in the source registry. These indicators let a model tell the difference between a measured value and an imputed value.

The current training vector has 33 values: 18 model properties and 15 missing-value indicators.

Target Score

target_score is copied from the source dataset score. The build process does not create a new score with another language model and does not apply a final score transformation. The released values currently range from 0.0 to 1.0.

The meaning of the score depends on score_type:

• routerbench_score: RouterBench's model evaluation score
• llm_judge_correctness: SPROUT's judge-based correctness score
• benchmark_score: LLMRouterBench's benchmark score

These scores share the same numeric range, but they are not guaranteed to measure exactly the same behavior. Users should inspect source and score_type when doing detailed analysis.

Recommended Split

Do not randomly split individual rows. A prompt usually appears once for many candidate models, so splitting rows can place the same prompt in both training and validation data.

The included loader splits by prompt_id into approximately:

• 80% training prompts
• 10% validation prompts
• 10% evaluation prompts

The normalization statistics are fitted on the training split only.

Intended Use

This dataset is intended for:

• Training prompt-to-model routing models
• Studying quality, cost, latency, and capability-aware model selection
• Regression or ranking experiments over candidate AI models

It is not intended to be treated as a universal leaderboard of model quality. The source evaluations use different tasks, judges, serving providers, and measurement procedures.

Limitations

• Some model-vector fields are imputed when direct source data was unavailable.
• Imputed values are accompanied by missing-value indicators.
• Latency and throughput values may come from different providers or observed benchmark runs and are not perfectly comparable.
• Provider pricing and serving behavior can change over time.
• Model names and routes may represent different versions or serving routes of a similar base model.
• Source score definitions are not identical across the three datasets.

Reproducibility

The canonical dataset is built by:

python scripts/build_training_data.py

The source-to-canonical transformation and model registry are maintained in the Kairos repository. The raw source files are not required to use the released Parquet file.

License

The source datasets have their own licenses and terms. Users should review the license and usage terms of RouterBench, SPROUT, and LLMRouterBench before redistributing or using this derived dataset.