Intro — Abhishek | Contributing to the Neural Extraction Framework ahead of GSoC 2027

Projects GSOC (Google Summer of Code)
1

Hi all — I’m Abhishek. I’m starting to contribute to the Neural
Extraction Framework and plan to apply for GSoC 2027 with DBpedia.

I know it’s early and the 2027 ideas aren’t out yet — posting now
because I’d rather spend the next six months doing actual work on the
codebase than turn up in March with a proposal and no commits.

Where I’ve got to so far

I’ve read through the NEF 2.0 work from GSoC 2026 (GSoC26/) — the
final report, CHANGES.md, and the reproducibility bundle. A few
things I took away:

  • The metric is exact-triple set matching after underscore/space
    stripping, macro-averaged as the mean of per-domain means (not a
    global mean over sentences) — monument at 19 sentences and city
    at 217 weigh equally.
  • Single-run noise floor is ±0.03 F1, so I’ll treat anything inside
    that as a match rather than a discrepancy.
  • The two biggest v14 wins (the cross-domain literal-flag leak, and
    bare-title sense resolution) came out of offline autopsies at zero
    LLM cost, which seems like the cheapest place for a newcomer to be
    useful.

What I’m planning next

Reproducing one domain from Text2KGBench dbpedia_webnlg end to end —
probably 3_airport — and posting the number next to the published
0.6051, with my setup and any deviations, whichever way it comes out.

Two questions

  1. REPORT.md §7 says the data snapshots (index ~1.4 GB, store ~15 GB)
    are planned for publication on the Databus. Are those up anywhere
    yet? Rebuilding the 16.2M-mention index and loading ~18 GB into
    oxigraph is the main barrier to reproducing the benchmark, and I
    suspect it’s what stops most people who try.

  2. The GSoC26/README.md future-work section lists four directions —
    anonymous-node grounding, Wikipedia-scale extraction,
    wikidata_tekgen generalisation, and the model-scaling ablation.
    Is there one mentors would most like picked up? Happy to start
    groundwork on whichever is most useful rather than guessing.

Also happy to take on smaller things in the meantime — there’s no CI
in the repo right now, and the two offline test suites
(test_date_style.py, test_gate_patches.py) would run in Actions
as-is if that’d be welcome.

Thanks!
Abhishek