amkdg/chatbot-arena-conversations-Embeddings

Chatbot Arena Conversations Embeddings

[image]

Embeddings of agie-ai/lmsys-chatbot_arena_conversations, produced with amkdg/Qwen3-Embedding-8B-NVFP4 — 4096-d, L2-normalized float16 (cosine = dot product).

• —65,960 conversations → 65,960 vectors
• —emb.npy — float16 [65960, 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")           # [65960, 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 agie-ai/lmsys-chatbot_arena_conversations — sourceref is `<questionid>|<a|b>` — arm a or b of that battle.

Notes

Both arms of 32,980 original LMArena human-preference battles (2023, ~20 models, real users with human votes) -> 65,960 conversations. tag = the model that produced the arm. Pulled from the ungated agie-ai/lmsys-chatbot_arena_conversations mirror of the gated original; 19 duplicate battles removed. Distinct from arena-human-preference-140K, arena-expert-5K, and search-arena-24K.