| license: apache-2.0 | |
| language: [en, ja, ko, multilingual] | |
| library_name: infonex | |
| tags: [information-extraction, infon, ner, relation-extraction, polarity, automotive, multilingual, gliner, fp16] | |
| # infon-extract (fp16) | |
| Grounded **infon** extractor: paragraph -> typed, grounded, polarity-aware spans/relations. | |
| Fine-tuned from `fastino/gliner2-multi-v1` (real span_rep + classifier weights, inherited | |
| calibration) on a 50k EN/JA/KO automotive corpus, + a 3-way polarity head. | |
| **Key fix vs base GLiNER2:** subword splitter + char-offset grounding -> CJK grounds 100% | |
| (base GLiNER2 tags whole JA sentences as one span). Partial-freeze fine-tune (emb+layers0-5 | |
| frozen) so calibration survives. GATE: 100% grounded EN/JA/KO, fp16 lossless vs fp32. | |
| Files: `model_fp16.safetensors` (encoder+span_rep+classifier+count), `polarity_head_fp16.pt`, | |
| `infonex_config.json` (arch), tokenizer/config. Use via the `infonex` package. | |
| Precision: fp16 (588MB). Runs on EC2 GPU (native) and Lambda CPU (upcast). | |