beatless-ai/polaris-ccf

Polaris CCF AI Papers

A curated, cross-validated, and richly enriched dataset of 155,652 papers from the 11 top AI conferences classified by CCF, spanning 1982–2026. Built by the Polaris pipeline.

Papers	155,652
CCF-A strict	AAAI · NeurIPS · ACL · CVPR · ICCV · ICML · IJCAI (7)
High-value extended	ICLR (unranked) · ECCV (CCF-B) · EMNLP (CCF-B) · CHI (CCF-A, HCI) (4)
Years	1982–2026
Authors	169,407 (deduplicated)
Author links	712,010
URL links	1,029,924
Abstract coverage	92.2%
Embedding model	all-MiniLM-L6-v2 (384-dim)
Knowledge-graph triplets	59,121 across 80,807 papers (51.9%)

Files

The dataset preserves the pipeline's directory structure. Download with --local-dir outputs/ai_ccf_a/ to drop straight into the Polaris repo's expected layout.

File	Size	Description
tables/ai_ccf_a_db_export_v2.csv	380 MB	Main table. One row per paper, 50 columns.
tables/ai_authors.csv	11 MB	Deduplicated author table (169 K rows).
tables/ai_paper_authors.csv	56 MB	Many-to-many paper↔author with author order (712 K rows).
tables/ai_paper_links.csv	225 MB	Per-paper URLs across 9 link types (1.03 M rows).
tables/ai_audit_report.md	3 KB	Data-quality audit summary.
knowledge/ai_paper_chunks.jsonl	254 MB	Per-paper text chunks (title + authors + abstract + venue). 1 line per paper.
knowledge/ai_paper_embeddings.npy	229 MB	Dense embeddings (155652, 384), aligned row-by-row with ai_paper_chunks.jsonl.
knowledge/ai_embedding_id_map.json	4.1 MB	Maps embedding row → paper_id.
knowledge/ai_knowledge_manifest.json	1 KB	Build metadata for the embedding index.
kg/knowledge_graph.json	16 MB	Method × dataset × task triplets extracted from titles + abstracts.
kg/kg_summary.json	3 KB	Triplet counts and entity coverage.
trends/trend_analysis.json	85 KB	50-topic clustering with hot/emerging/declining classification.
trends/trend_report.md	6 KB	Human-readable trend narrative.
trends/topic_centers.npy	76 KB	Cluster centroids (50, 384).
trends/topic_labels.npy	609 KB	Per-paper topic label (155652,).

Quick load

As a DataFrame (papers table)

import pandas as pd
df = pd.read_csv("tables/ai_ccf_a_db_export_v2.csv")
print(df.shape)            # (155652, 50)
print(df["conference"].value_counts())

As a HuggingFace Dataset

from datasets import load_dataset
ds = load_dataset("20bytes/polaris-ccf", "papers", split="train")

Semantic search with embeddings

import numpy as np, json
from sentence_transformers import SentenceTransformer

emb = np.load("knowledge/ai_paper_embeddings.npy")          # (155652, 384)
chunks = [json.loads(l) for l in open("knowledge/ai_paper_chunks.jsonl")]
model = SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2")

q = model.encode("retrieval augmented generation for code")
scores = emb @ q / (np.linalg.norm(emb, axis=1) * np.linalg.norm(q))
top = scores.argsort()[::-1][:10]
for i in top:
    print(scores[i], chunks[i]["title"])

Main-table schema (50 columns)

Identity & provenance

• paper_id, canonical_paper_id — primary key and dedup anchor
• dblp_key, identity_status, identity_match_method, identity_confidence
• source_host, source_sites_json, matched_both_sites

Bibliographic

• conference, year, title, normalized_title
• authors_text, authors_json, author_count
• group_name_raw, presentation_type, track_name, paper_type
• is_main_track, is_ccf_countable_paper, group_confidence, main_track_status

Links

• openreview_url, arxiv_url, paper_url, pdf_url, doi
• best_url, best_pdf_url

Content & enrichment

• abstract, abstract_source
• semantic_scholar_paper_id, s2_corpus_id, s2_url, s2_status
• s2_match_method, s2_match_confidence
• citation_count, influential_citation_count, reference_count
• fields_of_study_json, tldr

CCF classification

• ccf_rank, ccf_category, ccf_area_name
• ccf_official_name, ccf_official_url, scope

How it was built

The pipeline lives in the Polaris repository. Roughly:

1. Collect from open conference proceedings (CVF Open Access · ACL Anthology · OpenReview · PMLR) with title-level cross-validation.
2. Anchor to DBLP for canonical identity; resolve duplicates via DOI / arXiv ID / OpenReview ID / title+first-author.
3. Enrich with Semantic Scholar (citations, TLDR, fields of study), OpenAlex (backup abstracts), and arXiv (PDF links).
4. Normalize track names and groups using a 40-rule mapping table.
5. Embed each paper (title + authors + venue + abstract) with all-MiniLM-L6-v2.
6. Cluster into 50 topics with MiniBatchKMeans and classify trends by year.
7. Extract method × dataset × task triplets via pattern matching.

Sources & attribution

Source	License	What it contributed
DBLP	CC0 (Public Domain)	Canonical identity keys, venue series
Semantic Scholar API	Subject to S2 API Terms	Citation counts, TLDRs, fields of study
OpenAlex	CC0	Backup abstracts and metadata
arXiv	arXiv ToS	PDF links and identifiers
CCF	Cited as metadata	Venue tier rankings (CCF-A)
CVF Open Access · ACL Anthology · OpenReview · PMLR	Public conference pages	Title/author metadata

The dataset combines metadata from these sources. Abstracts come from official conference proceedings where available, with OpenAlex as backup. No full-text PDFs are redistributed.

License

Released under CC-BY-4.0. You may use, share, and adapt this dataset with attribution.

When citing or using this dataset, please also acknowledge the upstream sources above per their respective licenses.

Citation

@misc{polaris_ccf_ai_2026,
  title  = {Polaris CCF AI Papers: a cross-validated dataset of 155K papers from 11 top AI conferences in CCF rankings},
  year   = {2026},
  url    = {https://huggingface.co/datasets/20bytes/polaris-ccf},
  note   = {Pipeline: https://github.com/20bytes/Polaris}
}

Limitations & notes

• Abstract gaps: ~7.8% of papers have no abstract (mostly pre-2010 entries or workshop/demo tracks not covered by S2/OpenAlex).
• Citation freshness: citation counts are a snapshot from May 2026 via Semantic Scholar; treat them as a relative signal, not absolute truth.
• Author disambiguation: based on normalized names. Same-named authors may collide; cross-check with s2_author_id where available.
• Knowledge graph: extracted via pattern matching over a hand-curated entity vocabulary (~200 methods × 100 datasets × 80 tasks). High precision, recall varies by subfield.
• Year coverage: dense from 2010 onward; pre-2010 coverage is best-effort and varies by venue.

Changelog

• v0.1.0 (2026-05): Initial release covering 11 venues, 1982–2026.