amkdg/bluemoon-Embeddings

Bluemoon Roleplay Embeddings

Embeddings of Squish42/bluemoon-fandom-1-1-rp-cleaned, produced with amkdg/Qwen3-Embedding-8B-NVFP4 — 4096-d, L2-normalized float16 (cosine = dot product).

• 3,338 conversations → 9,890 vectors
• emb.npy — float16 [9890, 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")           # [9890, 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 Squish42/bluemoon-fandom-1-1-rp-cleaned — source_ref is the thread id in data/full/bluemoon.train.json.

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

Notes

Long-form human↔human 1-on-1 roleplay threads from the Bluemoon forum (the biggest cleaned split, 3,338 threads). Both participants are human; mapped to user/assistant for a consistent two-role transcript. Not filtered for NSFW content.