| --- |
| license: mit |
| library_name: onnx |
| tags: |
| - token-classification |
| - ner |
| - resume |
| - distillation |
| base_model: microsoft/deberta-v3-xsmall |
| --- |
| |
| # onnx-ner-student |
|
|
| DeBERTa-v3-xsmall BIO token classifier distilled from the production **GLiNER2** |
| teacher (`fastino/gliner2-base-v1`) for résumé entity extraction. Replaces the |
| ~1 GB GLiNER2 runtime in curriculo-ai to fit the t3.medium memory budget. |
|
|
| **16 entity types**, each an independent BIO sequence (a token may be B for |
| several types at once — e.g. `CI/CD` is both `technical_skill` and `framework`). |
|
|
| ## Files |
| - `model.onnx` — FP32 |
| - `model_quantized.onnx` — INT8 dynamic (runtime default) |
| - `labels.json` — the 16 type names, index-aligned to the output head |
| - tokenizer files (fast/`tokenizers`-loadable, torch-free) |
|
|
| ## I/O |
| `input_ids`, `attention_mask` `[B, T]` → `logits` `[B, T, 16, 3]` |
| (argmax over the last dim → per-type BIO tag `0=O,1=B,2=I`; decode with |
| `ner_dataset.decode_spans`). |
|
|
| ## Results (held-out test vs teacher, best `ct0.5_lr2e-04_ep24`) |
| micro-F1 **0.8924**, precision 0.8881, recall 0.8968. |
|
|
| | type | F1 | |
| |------|----| |
| | award | 0.9474 | |
| | certification | 0.5689 | |
| | degree | 0.9064 | |
| | field_of_study | 0.8165 | |
| | framework | 0.7987 | |
| | industry | 0.8319 | |
| | interest | 0.9818 | |
| | job_title | 0.932 | |
| | language | 0.9515 | |
| | location | 0.8925 | |
| | organization | 0.9726 | |
| | person_name | 0.9655 | |
| | soft_skill | 0.8108 | |
| | technical_skill | 0.8536 | |
| | technology | 0.92 | |
| | tool | 0.8772 | |
|
|
