cstorm125/graphiv_title_abstract_embeddings

GraphIV title and abstract embeddings

Every arXiv paper in GraphIV, an open citation graph of arXiv, embedded by seven open models from 0.6B to 27B parameters, made to measure how much a larger embedding model moves GraphIV's retrieval. Each model embeds the same document GraphIV embeds, Title: ...\nAbstract: ..., cut at 512 tokens, after its own document prompt.

Source

Every paper is one of the 3,186,931 in kkuroma/GraphIV at tag GraphIV-2026-10-01, and arxiv_id joins these vectors to its papers table, whose embedding column holds GraphIV's own nomic-embed-text-v2-moe vectors. manifest.json lists each file's sha256.

Models

config	model	dims	document prompt	query prompt
harrier-0.6b	microsoft/harrier-oss-v1-0.6b	1024	none	the instruction below
harrier-27b	microsoft/harrier-oss-v1-27b	5376	none	the instruction below
nemotron-1b	nvidia/Nemotron-3-Embed-1B-BF16	2048	passage:	query:
nemotron-8b	nvidia/Nemotron-3-Embed-8B-BF16	4096	passage:	query:
qwen3-0.6b	Qwen/Qwen3-Embedding-0.6B	1024	none	the instruction below
qwen3-4b	Qwen/Qwen3-Embedding-4B	2560	none	the instruction below
qwen3-8b	Qwen/Qwen3-Embedding-8B	4096	none	the instruction below

The instruction is Instruct: Given a research question, retrieve the arXiv papers a literature review answering it would cite\nQuery: . Every model ran through vLLM's pooling runner in bfloat16, and its vectors are L2-normalized and stored as float16.

Schema

arxiv_id (string) and embedding (float16 list, the model's width), one row a paper, sorted by arxiv_id, in row groups of 50,000.

Use

from datasets import load_dataset

vectors = load_dataset("cstorm125/graphiv_title_abstract_embeddings", "harrier-0.6b", split="train")

Queries must carry the model's own query prompt, and every vector is unit length, so a dot product is the cosine.

Results

Test split of kkuroma/GraphIV_benchmarks, recall@1000 (ConsensusMatch at 100), with BoundedField's shipped recipes, against GraphIV's own nomic-embed-text-v2-moe vectors:

model	SurveyMatch, cosine only	SurveyMatch, survey	FrontierMatch, frontier	ConsensusMatch, consensus
nomic-v2-moe (GraphIV's)	0.282	0.766	0.745	0.787
nemotron-1b	0.299	0.784	0.741	0.803
qwen3-0.6b	0.322	0.774	0.738	0.828
harrier-0.6b	0.342	0.787	0.778	0.819
qwen3-4b	0.387	0.809	0.787	0.833
qwen3-8b	0.409	0.813	0.794	0.838
nemotron-8b	0.409	0.827	0.788	0.845
harrier-27b	0.432	0.815	0.792	0.844

Cosine alone gains most from size, while with the citation recipes the gain stops growing past 8B. The full tables, with PrecisionMatch, refit recipes and 95% intervals, are in scripts/embedding_sweep/README.md of the GraphIV repository.

License

The vectors are derived from arXiv titles and abstracts through kkuroma/GraphIV, which is CC BY-NC 4.0, so this dataset is CC BY-NC 4.0 too. The embedding models carry their own licenses, linked above.