amkdg/collectivecognition-Embeddings

CollectiveCognition Embeddings

Embeddings of the CollectiveCognition chats-data snapshots (2023-09-22, 2023-09-27, 2023-10-16), produced with amkdg/Qwen3-Embedding-8B-NVFP4 — 4096-d, L2-normalized float16 (cosine = dot product).

• 343 conversations → 391 vectors
• emb.npy — float16 [391, 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")           # [391, 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 the CollectiveCognition chats-data snapshots (2023-09-22, 2023-09-27, 2023-10-16) — source_ref is the chat title.

Notes

All three published snapshots merged and deduplicated by title (longest version kept) → 343 unique shared-ChatGPT conversations. tag = ChatGPT model.