amkdg/lmsys-chat-1M-Embeddings

LMSYS-Chat-1M Embeddings

Embeddings of AarushSah/lmsys-chat-1m, produced with amkdg/Qwen3-Embedding-8B-NVFP4 — 4096-d, L2-normalized float16 (cosine = dot product).

• 1,000,000 conversations → 1,004,674 vectors
• emb.npy — float16 [1004674, 4096]
• meta.parquet — one row per vector, aligned with emb.npy: id, uuid, 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")           # [1004674, 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 AarushSah/lmsys-chat-1m — source_ref is the upstream conversation_id.

ref = meta.iloc[0].source_ref      # -> the matching conversation in the source dataset

Notes

The full corpus — no dedup and no flagged/toxic removal. count is always 1.