rizzo-pii-0.3B-onnx

Rizzo PII 0.3B ONNX

Two versions:

• v1 (repo root): an unquantized ONNX export of rizzoaiacademy/rizzo-pii-0.3B.
• v2 (v2/): a fine-tune of v1 on Italian spoken dialogues. It keeps v1's results on the original benchmarks.

v2: fine-tuned on Italian dialogues

Training:

• Starting point: a full fine-tune of rizzo-pii-0.3B.
• Data, mixed one to one:
  • about 5,000 synthetic Italian spoken dialogues written by an LLM;
  • a replay of rizzo-pii-it-dataset, so the legal domain is not forgotten.
• Settings: 2 epochs, learning rate 2e-5, 1,024-token windows.
• Checkpoint: chosen on held-out dialogues and the dataset's dev split together.

Labels, tokenizer and inputs are the same as v1's, so v2 is a drop-in replacement. v2/model-fp16.onnx is fp16 (616 MB), and the other files in v2/ are copies of the root ones.

In the tables, each cell shows the share of entities left unmasked first, and the share of non-PII words masked by mistake after the slash.

Original benchmarks: nothing lost. These are the validation sets of rizzo-pii-it-dataset, scored with rizzo-pii's own protocol (word tokens, strict type and span match).

In validation_real, 64% of the rows carry WordPiece ## pieces. The second row is the same text joined back into words, which is how real text arrives.

Italian dialogues. 998 held-out synthetic dialogues with 22,474 entities, written by a different LLM than the training dialogues. Each line is classified and every detected value is masked wherever it repeats. A regex layer adds emails, phones and IDs, and every model gets the same post-processing.

Most of the gap is names: v1 misses 8,224 of the 19,565 names, v2 misses 74.

v1: export of rizzo-pii-0.3B

Unquantized FP32 ONNX export of rizzoaiacademy/rizzo-pii-0.3B at revision d4f989333bcda1fcb188785a12937b0f52b52580.

Exported with Optimum 1.24.0 at opset 17. Two weight variants are provided:

• model-fp16.onnx: 615,826,010 bytes; recommended smaller CPU artifact.
• model.onnx: 1,230,827,904 bytes; canonical FP32 reference.

FP16 was evaluated on all 7,000 examples and 255,154 word tokens in the released validation_real.jsonl. It produced 100% token-argmax and exact entity-span agreement with FP32. Both variants achieved micro-F1 0.9886719380. Full results are in validation-fp16.json.

Inputs: input_ids, attention_mask (int64). Output: token-classification logits. config.json contains the BIO label map; tokenizer.json provides tokenization and character offsets.

Exact build versions, source revision, file hashes, and parity tolerances are recorded in manifest.json. For model details, intended use, metrics, and licensing, refer to the original artifact page.