amkdg/oasst2-Embeddings

OpenAssistant oasst2 Embeddings

Embeddings of OpenAssistant/oasst2, produced with amkdg/Qwen3-Embedding-8B-NVFP4 — 4096-d, L2-normalized float16 (cosine = dot product).

• 128,848 conversations → 128,850 vectors
• emb.npy — float16 [128850, 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")           # [128850, 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 OpenAssistant/oasst2 — source_ref is <message_tree_id>:<leaf message_id>: the root-to-leaf path ending at that message.

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

Notes

Each root-to-leaf branch of every message tree is one conversation (128,848 from 70,642 trees). tag is the conversation language.