arxiv-funding-entity-extractions

Funder/award entity extractions over cometadata/arxiv-funding-statements.

• Extractor: funding-entity-extractor (vLLM + LoRA)
• Base model: meta-llama/Llama-3.1-8B-Instruct
• LoRA: cometadata/funding-extraction-llama-3.1-8b-instruct-artifact-data-mix-grpo-mixed-reward
• Hardware: A100-large bf16, concurrency 256
• Total rows: 1,823,650

Configs

• predictions (default) — original extractions, no ROR enrichment.
• predictions_with_ror — same rows, with ROR IDs attached to each funder via the marple match service plus a curated assignment layer (see ROR enrichment below).

Schema

Per row in predictions/*.parquet — every input column from cometadata/arxiv-funding-statements (arxiv_id, shard_id, doc_id, input_file, row_idx, predicted_statements, predicted_details, text_length, latency_ms, error) plus four extraction columns:

• extracted_funders: list<list<struct{funder_name, awards: list<struct{award_ids, funding_scheme, award_title}>}>> Outer list parallels predicted_statements; inner list is the funders found in that statement (null entry on parse error).
• extraction_raw: list<string> — verbatim model output per statement.
• extraction_error: list<string> — null on success, else "ParseError: ..." or "HTTPError: ...".
• extraction_latency_ms: list<float64> — per-statement wall-clock latency.

ROR enrichment

predictions_with_ror/*.parquet is the same 184 shards with two additional fields appended to the leaf funder struct in extracted_funders:

• ror_id: nullable string — matched ROR URL (https://ror.org/<id>) or null when no match was found.
• match_confidence: nullable double — confidence score from the match service, or null for curated assignments.

Enriched leaf struct shape:

struct{
  funder_name: string,
  awards: list<struct{award_ids, funding_scheme, award_title}>,
  ror_id: string?,
  match_confidence: double?
}

Pipeline

1. Unique funder names extracted from predictions/*.parquet.
2. Matched against the marple match service running the funder strategy authored by Jason Portenoy (a modified single-search strategy originally by Dominika Tkaczyk).
3. A curated assignment layer fills in 97 high-volume funder names that the match service rejects (3-letter acronyms like NSF, NIH, DOE; ambiguous 4-letter acronyms; common full-name variants). Curated entries write match_confidence: null so they are distinguishable from match-service rows.
4. Reconciliation joins matches back to the original parquets shard by shard, preserving every other column verbatim.

Coverage

• Total funder occurrences across the corpus: 4,827,173
• Matched (marple service only): 2,035,930 (42.18%)
• Matched (marple + curated assignments): 3,189,332 (66.07%)