amkdg/search-arena-24K-Embeddings

LMArena Search-Arena-24K Embeddings

Embeddings of lmarena-ai/search-arena-24k, produced with amkdg/Qwen3-Embedding-8B-NVFP4 — 4096-d, L2-normalized float16 (cosine = dot product).

• 48,138 conversations → 48,694 vectors
• emb.npy — float16 [48694, 4096]
• meta.parquet — one row per vector, aligned with emb.npy: id, uuid, tag, chunk, n_chunks, count, source_ref
• manifest.json — counts and provenance

Usage

import numpy as np, pyarrow.parquet as pq

emb  = np.load("emb.npy", mmap_mode="r")           # [48694, 4096] float16
meta = pq.read_table("meta.parquet").to_pandas()   # one row per vector, aligned with emb

# A conversation = consecutive rows sharing one `uuid` (`chunk == 0` marks its start);
# conversations longer than 8192 tokens span several chunk-rows.
starts = meta.index[meta.chunk == 0]               # first row of each conversation

Source mapping

Each row carries source_ref, the locator back into lmarena-ai/search-arena-24k — source_ref is the upstream conv_id (from system_{a,b}_metadata).

Notes

Both arms of 24,069 LMArena search battles (web-search-augmented models) → 48,138 conversations. tag = the model that produced the arm.