carsondial/qwen-8b-embed

Qwen3-Embedding-8B @ 1024d — full PyNIFE distillation corpus

Pre-computed teacher embeddings across 15 source datasets spanning documents, queries, and symmetric sentence pairs. Designed to replicate the full PyNIFE / LEAF two-stage training recipe against Qwen/Qwen3-Embedding-8B as the teacher.

Headline numbers

• Total rows: 11,488,955
• Total tokens embedded: 1469.8M
• Total cost: ~$146.98 on Fireworks serverless at $0.10/1M tokens
• Embedding dim: 1024 (MRL-native; truncate to 256/512 as needed downstream)
• Max input tokens: 2048 (client-side truncation via Qwen3 tokenizer)
• Normalization: L2-normalized unit vectors

Schema (all configs identical)

column	type
text	string (possibly truncated to ≤2048 tokens)
embedding	float32[1024], unit-norm
role	"doc" | "query" | "symmetric"

Per-source breakdown

config	upstream	role	rows	tokens
english-words-definitions	MongoDB/english-words-definitions	doc	466,357	15.6M
fineweb	HuggingFaceFW/fineweb	doc	2,100,000	1204.8M
gooaq	sentence-transformers/gooaq	query	3,012,496	29.8M
miracl	sentence-transformers/miracl	query	2,863	0.0M
lotte	mteb/lotte	query	13,028	0.2M
snli	stanfordnlp/snli	symmetric	629,334	6.4M
paws	google-research-datasets/paws	symmetric	1,291,304	34.5M
squad	sentence-transformers/squad	query	87,599	1.1M
mldr	sentence-transformers/mldr	doc	10,000	0.1M
msmarco	sentence-transformers/msmarco-corpus	query	1,010,916	7.6M
msmarco_docs	sentence-transformers/msmarco-corpus	doc	2,000,000	154.9M
PubMedQA	qiaojin/PubMedQA	query	272,518	6.2M
swim-ir-monolingual	nthakur/swim-ir-monolingual	query	501,371	6.9M
trivia_qa	mandarjoshi/trivia_qa	query	87,622	1.7M
mr-tydi	sentence-transformers/mr-tydi	query	3,547	0.0M

Two-stage training recipe (following LEAF / PyNIFE)

Interleaving docs and queries during distillation does not work well (see Tulkens' README). The recommended recipe is:

1. Pretrain on doc-like sources: concatenate the configs where role == "doc" (msmarco_docs, mldr, fineweb, english-words-definitions).
2. Finetune with a lower learning rate on query-like sources: concatenate the configs where role == "query" (msmarco, gooaq, squad, swim-ir-monolingual, trivia_qa, PubMedQA, miracl, mr-tydi, lotte).

The symmetric sources (snli, paws) are sentence-pair corpora useful for STS-style alignment; use at your discretion.

from datasets import load_dataset, concatenate_datasets

REPO = "REPLACE_WITH_HF_REPO_ID"

# Stage 1: documents
doc_configs = ["msmarco_docs", "mldr", "fineweb", "english-words-definitions"]
doc_train = concatenate_datasets([
    load_dataset(REPO, c, split="train") for c in doc_configs
])

# Stage 2: queries
query_configs = ["msmarco", "gooaq", "squad", "swim-ir-monolingual",
                 "trivia_qa", "PubMedQA", "miracl", "mr-tydi", "lotte"]
query_train = concatenate_datasets([
    load_dataset(REPO, c, split="train") for c in query_configs
])

Why no instruction prompt?

Per PyNIFE's empirical finding: static models cannot use instructions meaningfully, because with no cross-token interaction the instruction prompt can only produce a constant offset in embedding space — invisible to cosine similarity ranking. So teacher embeddings here are computed on plain text.

Asymmetric retrieval pattern

This corpus is raw material for an asymmetric architecture: expensive teacher for document indexing, cheap distilled student for online queries. See PyNIFE and LEAF for the theory.

Reproducibility

Generated by build_corpus.py. Deterministic within a given set of upstream dataset snapshots. Fireworks accounts/fireworks/models/qwen3-embedding-8b with dimensions=1024; vectors re-normalized L2 client-side after receipt (MRL truncation returns non-unit vectors).