amkdg/bing-conversations-Embeddings

Bing / Sydney Conversations Embeddings

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

• 644 conversations → 650 vectors
• emb.npy — float16 [650, 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")           # [650, 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 david-ar/bing-chat-logs — source_ref is the original tweet/reddit post URL the screenshot was captured from.

Notes

Real Bing Chat ("Sydney") transcripts — images scraped from public Twitter/Reddit posts (Feb–Apr 2023), transcribed from the screenshots. Non-conversation/image-only stubs dropped (644 of 660). tag = capture platform.