amkdg/gpt-roleplay-realm-Embeddings

GPT Roleplay Realm Embeddings

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

• 8,700 conversations → 8,700 vectors
• emb.npy — float16 [8700, 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")           # [8700, 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 IlyaGusev/gpt_roleplay_realm — source_ref is <char_id>:<dialogue_index> (0–19) into the character's dialogues.

Notes

Synthetic: humans wrote only seed lists — GPT-4 generated the characters/topics and the 1st of each character's 20 dialogues, GPT-3.5 the other 19. en (216) + ru (219) are independently generated, not translations. Character persona is prepended as a system turn. tag = language.