amkdg/arena-hard-auto-Embeddings

Arena-Hard-Auto Embeddings

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

• 1,250 prompts → 1,251 vectors
• emb.npy — float16 [1251, 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")           # [1251, 4096] float16
meta = pq.read_table("meta.parquet").to_pandas()   # one row per vector, aligned with emb
starts = meta.index[meta.chunk == 0]               # first row of each conversation

Source mapping

Each row carries source_ref, the locator back into lmarena-ai/arena-hard-auto — source_ref is <version>|<uid> into that version's question.jsonl.

Notes

Both Arena-Hard-Auto benchmark prompt sets, embedded as the union: v0.1 (500 prompts, 2024) + v2.0 (750 prompts, 2025) = 1,250. The two sets are disjoint (zero shared prompts); v2.0 is the newer, harder re-curation. User prompts only, no model answers. tag = arena-hard-v0.1 / arena-hard-v2.0.