von-ft

Fine-tune of Von 1.3 (https://huggingface.co/wfzyx/von, ModernBERT-large backbone + Option-Marker scoring head, 395M parameters) on a custom decision corpus. Same architecture and wire protocol as Von; install and usage are in https://github.com/wfzyx/von.

• Base checkpoint: Von 1.3
• Training: 1 epoch, 12,917 steps, independent_options attention and digit-split tokenisation
• Validation: 75.89% accuracy (loss 1.6333) on the held-out data_ft validation split
• License: Apache-2.0 (inherited from ModernBERT and Von)

Training accuracy reached ~100% while validation is 75.89%, so expect a generalisation gap outside the fine-tuning distribution.

JevBench

suite	n	von-1.2	von-ft	Δ	p
jev_standard	72	55.6	66.7	+11.1pp	0.134
jev_hard	111	37.8	32.4	−5.4pp	0.286
jev_easy	48	100	100	0	1.0
jabr_v2	869	71.3	70.7	−0.7pp	0.689
pooled (all 1100)	1100	68.2	67.8	−0.4pp	0.832

Files

File	Purpose
model.safetensors	Fine-tuned ModernBERT-large backbone
option_marker.pt	Option-Marker scoring head
marker_calibration.json	Calibration metadata and independent_options: true. The SDK reads this flag; do not drop it.
config.json, tokenizer*	Standard HF config and tokenizer

Usage

pip install von-sdk
$env:VON_CHECKPOINT_DIR="D:\von-ft"; uv run von serve --host 0.0.0.0 --port 8001