amkdg/sharegpt-Embeddings

ShareGPT Embeddings

Embeddings of RyokoAI/ShareGPT52K (90,665 conversations, despite the “52K” in the name), produced with amkdg/Qwen3-Embedding-8B-NVFP4 — 4096-d, L2-normalized float16 (cosine = dot product).

• 90,665 conversations → 119,527 vectors
• emb.npy — float16 [119527, 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")           # [119527, 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 RyokoAI/ShareGPT52K — source_ref is the upstream conversation id.

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