Downloads · 30 days
163
40% of all-time downloads
guard-ft-v19-trimmed-int8
gauthierrobert2/guard-ft-v19-trimmed-int8
guard-ft-v19-trimmed-int8 is a token classification model from gauthierrobert2. Use it when you need labels on individual words, such as names.
GUARD-FT round 19 (AgentDocs' EN+FR legal-document PII token classifier), vocabulary-trimmed (250,002 - 53,751 pieces) then dynamically quantized to int8 (onnxruntime.quantization.quantizedynamic, QInt8, no calibratio…
All-time downloads
407
Public
Repo size
360 MB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.onnx360 MB · 99%
At a glance
- Task
- Token Classification
- Model type
- xlm-roberta
- Access
- Public
- Created
- Aug 26, 2026
- Updated
- Aug 26, 2026
- SHA
- 6c2f4a76
Try a prompt
Base models
Model card
From the Hugging Face model README
guard-ft-v19-trimmed-int8
GUARD-FT round 19 (AgentDocs' EN+FR legal-document PII token classifier), vocabulary-trimmed
(250,002 -> 53,751 pieces) then dynamically quantized to int8
(onnxruntime.quantization.quantize_dynamic, QInt8, no calibration set — every MatMul weight
tensor; activations quantized per inference).
A real precision/speed trade-off, not a free size win (unlike the fp16 sibling, which is
numerically near-identical to fp32). Measured against gauthierrobert2/guard-ft-v19-trimmed across
a 259-document evaluation corpus:
| fp32 (trimmed) | int8 (this) | |
|---|---|---|
| stage-3 leaked | 234 | 243 (+9) |
| stage-3 ORG recall | 89.68% | 88.56% |
| stage-3 OTHER recall | 81.56% | 80.50% |
| census over-masked | 798 | 833 |
| census precision | 80.94% | 80.44% |
CPU is where it wins: ~1.9-2.9x faster than fp32 (measured, oneDNN/MLAS int8 kernels), 360 MB
against fp32's 1.4 GB. CUDA is where it loses badly: ~9-10x SLOWER than fp32 in a real detection
pass — dynamic quantization targets CPU kernels and has no efficient GPU execution path here. Loads
correctly under ONNX Runtime 1.24.1's CPU execution provider at ORT_ENABLE_ALL with no fusion
pre-pass needed (unlike fp16, which does).
- Weights: ~1.4 GB (fp32) -> ~360 MB (int8)
- Labels: PERSON, ORG, PUBLIC_BODY, ROLE, ADDRESS, PLACE, DOB, CONTACT (BIO scheme)
- Task
- Token Classification
- Type
- xlm-roberta
- Created
- Aug 26, 2026
- Updated
- Aug 26, 2026