amkdg/aya-dataset-Embeddings

Aya Dataset Embeddings

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

• 190,385 conversations → 191,650 vectors
• emb.npy — float16 [191650, 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")           # [191650, 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 CohereForAI/aya_dataset — source_ref is the conversation content-hash (= id); tag carries the language.

Notes

~190k human-written prompt/completion pairs authored by fluent-speaker annotators across 71 languages (Aya open-science initiative) — the human aya_dataset, NOT the templated + machine-translated 513M aya_collection. tag = language. Exact-duplicate pairs removed (11,970); aya has no per-row id, so source_ref is the content-hash.