RaghavendraSqwish/routerarena-with-embeddings

RouterArena with embeddings

This repository contains the complete-query portion of the legacy RouterArena 3x run r2_legacy_unconstrained_full_all_enabled_3x, together with aligned Qwen 3.5 prompt embeddings.

Recommended configuration

Use queries_with_embeddings for one row per prompt. It contains 8,345 rows and the following high-level fields:

• original query metadata and formatted prompt;
• the query-level oracle label and ranked six-model pool;
• six per-model aggregate summaries;
• embedding: a fixed-length list of 6,144 float32 values.

The other configurations retain the normalized RouterArena tables:

• query_oracle_labels: 8,345 query-level oracle rows;
• query_model_summary: 50,070 query/model aggregate rows;
• model_outcomes: 150,210 attempt-level outcome rows (three repeats for six models over every retained query).

Embedding provenance

The embedding vectors were generated with Qwen/Qwen3.5-2B at revision 15852e8c16360a2fea060d615a32b45270f8a8fc. They concatenate hidden states [-1, -9, -17] at the final input-token position, yielding 6,144 dimensions. Inputs use the v12_universal_short_15 analysis prompt, a 4,096-token context, and middle_out truncation.

The original compressed NumPy artifact is preserved at artifacts/qwen35_embeddings_8345x6144.npz. The exact analysis prompt, training-table report, model capability fingerprints, checksums, row counts, and alignment details are included under provenance/ and in manifest.json.

Alignment and completeness

The prompt/embedding join was validated before upload:

• embedding shape: (8345, 6144);
• all 8,345 global_index values are unique;
• the prompt, oracle, and model-summary ID sets match exactly;
• every query has exactly six model summaries;
• the outcome table contains all expected 150,210 rows.

This is the complete_queries_only coverage set. Fifty-five incomplete source queries are intentionally excluded, as recorded in the training-table report.