amkdg/toxic-chat-Embeddings

LMSYS Toxic-Chat Embeddings

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

• 10,165 conversations → 10,165 vectors
• emb.npy — float16 [10165, 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")           # [10165, 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 lmsys/toxic-chat — source_ref is the upstream conv_id.

Notes

Real user queries from the LMSYS Vicuna online demo, each with one model response, plus human toxicity / jailbreak annotations. tag = tox{0/1}_jb{0/1} (the human toxicity and jailbreak flags). 10,165 conversations, no duplicates.