hyper-gliner

Torch-only inference for grounded infon extraction + learned-sparse recall, multilingual (EN/JA/KO), on fine-tuned mDeBERTa + a BERT SPLADE retriever. No transformers runtime dependency — both encoders are reimplemented and parity-proven, so the package is small enough to ship in an AWS Lambda container image.

from hyper_gliner import InfonExtractor, SpladeRetriever

ex = InfonExtractor.from_pretrained("cp500/infon-extract")        # extraction + polarity + coref
infons = ex.extract("Samsung SDI supplies battery cells to BMW.")  # -> [Infon(...)]
#  Infon(Samsung SDI —supplies→ battery cells)  polarity=affirmed

# the {"anchors","infons","stats"} envelope the downstream pipeline consumes:
doc = ex.extract_doc("...", title="...")        # drop-in for extract_infons_neural

sp = SpladeRetriever.from_pretrained("cp500/infon-extract")        # loads the sparse/ subfolder
terms = sp.encode("electric SUV battery", mode="query")            # {term_id: weight}
doc = ex.extract_doc("...", sparse=sp)          # attaches per-infon splade_terms (recall tags)

What's in the box

All weights ship as fp16 in cp500/infon-extract (the sparse model under sparse/).

Dependencies

torch, tokenizers, safetensors, numpy. Optional huggingface_hub (only to download from the Hub; pass a local dir otherwise — e.g. weights baked into a Lambda image).

Design notes

• Torch-only, parity-proven. The mDeBERTa encoder (disentangled attention) and the BERT SPLADE encoder are reimplemented in pure torch and gated bit-close against the reference implementation (tests/test_parity.py 9.5e-5, tests/test_bert_parity.py 1e-5). The full decode is gated against the reference implementation (tests/test_decode_parity.py, exact on EN/JA/KO).
• CJK grounding. A subword splitter + char-offset grounding make Japanese/Korean ground at 100% (the base span-extraction architecture whitespace-splits and fails on spaceless CJK).
• Precision vs recall, separated. extract() applies an exact-span argmax filter (one canonical type per span, nesting preserved). The SPLADE head supplies the multi-label / implicit recall channel as splade_terms — feeding an inverted index, a SQL splade_terms side-table, and the term↔infon bipartite incidence a Sheaf GNN expands.
• Triple-level grounding filter (on by default). extract() / extract_doc() run a measured precision filter (decode_filter.filter_infons): it drops the predicates an adversarial grounding audit found ~95–100% non-grounded (supplies, partner, competes_with — the relation head fires them from the document theme, not the connecting clause), plus self-loops and non-entity (quantity/currency/spec) objects. On the audited set this lifts grounded precision 29.5%→58% and cuts hallucination 16.7%→6% while retaining 93.5% of truly-grounded infons. Pass filter_triples=False for raw, unfiltered output.
• fp16 only. int8-dynamic was measured to hurt the extraction model (grounding 14→5 spans, and larger on disk because the embedding table stays fp32); fp16 is the shipped artifact.

Consumer contract

extract_doc() returns the same {"anchors","infons","stats"} envelope, with each infon dict carrying subject / predicate / object / polarity(int) / mass{supports,refutes,theta} / span / confidence / relation_type / tense / anchors, so it is a drop-in for the downstream ingest pipeline + Memgraph writer + MCTS/IKL readers. See tests/test_consumer_contract.py.

License

Apache-2.0.