simoneschiavoi/VisionPsy-Nano-DomCal-EmbGuard

VisionPsy-Nano DomCal-EmbGuard

The flagship challenger of this release: 8W/5L/4T over 17 displayed rows vs QVAC q4_k_m-imat (same-harness re-run; MM-IFEval@2048 42.2 > 41.1) — the highest normalized score of all six ledger models (61.97) and the fastest observed in this 200-item solo sweep (786.7 ms/item).

A GGUF quantized derivative of QVAC's VisionPsy-Nano-460M built with P1 (domcal) × P4 (embguard) — domain-calibrated imatrix quantization with embedding-protected precision (input embedding tensor held at Q8_0) on a q4_k_m base. This is not a new trained checkpoint: the ~460M-parameter architecture is unchanged. Part of the VisionPsy-Nano release collection — see the Links section below.

Exact runtime pair

Component	File	Bytes
LM	visionpsy-nano-460m-q4_k_m-domcal-embguard.gguf	320862016
mmproj	mmproj-visionpsy-nano-460m-q8.gguf	108782144

Package size: 409.7 MiB. Use only this LM/mmproj mapping. SHA-256 checksums for both files ship in this repository as SHA256SUMS — verify after download with sha256sum -c SHA256SUMS.

Model at a glance

Base model	QVAC VisionPsy-Nano-460M (~460M parameters; SigLIP2 vision encoder + SmolLM2-360M backbone)
Techniques	P1 (domcal) — domain-calibrated imatrix quantization; P4 (embguard) — input embedding tensor (token_embd) held at Q8_0
Quantization	q4_k_m LM + q8 mmproj (QVAC's own projector artifact)
Calibration data	VQAv2-train + TextVQA-train, ChatML-formatted; training splits only — never benchmark test data, no evaluation images
Total size	409.7 MiB = 306.0 MiB LM + 103.7 MiB mmproj — +16.9 MiB (+4.3%) vs QVAC q4_k_m-imat (392.8)
Headline	8W/5L/4T over 17 displayed rows vs QVAC q4_k_m-imat (same-harness re-run); normalized 61.97, best of all six ledger models; best measured MMStar (47.13), MMBench (61.46), MathVista (48.60), MMVet (35.60); MM-IFEval@2048 42.24 > 41.11; 14W/3L/0T vs QVAC q4_0 (same-harness)
Observed speed	786.7 ms/item — fastest observed in this 200-item solo sweep, ahead of all 15 measured configurations (next: TriStack 795.4; q4_0 803.5; imat 813.2)

Why this build exists

The study's conclusion frames each package as a deliberate assignment of calibration and precision — and this is the maximum-quality assignment. Domain calibration repairs the importance information; embedding protection keeps the token-embedding tensor at Q8_0, where the study found quantization drift most expensive. The measured outcome: the reasoning capability area scores 52.3 — above even QVAC's fp32 published card (52.2) — from a package roughly 4× smaller than fp32. Mechanism honesty: the incremental EmbGuard step at q4_k_m precision was not separately significant against its domain-calibrated parent (+0.27 pooled point, p = 0.38), so the complete result is attributed to the configuration as shipped, not to embedding protection alone.

Vs the two rulers

The primary ruler is QVAC q4_k_m-imat, QVAC's flagship build — re-run in this same harness with a hash-pinned full-17 record (judged rows scored with a qwen3.6-27b API judge, a reconstruction of QVAC's judging protocol validated within ±1 pt of their published card on 6/8 judged benchmarks; QVAC's own card numbers remain labeled context, never medaled). QVAC q4_0 is the secondary same-harness ruler. Negative deltas = smaller/faster.

	QVAC q4_k_m-imat (same-harness re-run)	QVAC q4_0 (same-harness)
W/L/T over 17 displayed rows	8W/5L/4T	14W/3L/0T
Package size Δ (409.7 MiB)	+16.9 MiB (+4.3%)	+62.0 MiB
Speed Δ (786.7 ms/item)	−26.5 ms/item (−3.3%)	−16.8 ms/item

Measured results

Row-golds across the six-model ledger: MMStar 47.13, MMBench 61.46, MathVista 48.60, MMVet 35.60, AI2D 66.03, MM-IFEval@2048 42.24. The named losses stay visible: QVAC q4_k_m-imat keeps OCRBench (777 vs 766), DocVQA, ChartQA and InfoVQA — the document/OCR cluster stays with the imatrix build; QVAC q4_0 keeps RealWorldQA, MME and MMMU dev. The fastest-observed latency and the highest normalized score sit in the same package: the quality gains cost nothing in measured speed.

Limitations

• Exploratory scope: one seed (17), one harness, one GPU. Counts are descriptive rows, not universal-superiority claims; "fastest observed" is scoped to this 200-item solo sweep.
• Mechanism attribution: the EmbGuard increment was not separately significant against its calibrated parent (p = 0.38); the package is evaluated as a whole.
• Judged rows: qwen3.6-27b via OpenRouter — an attempted same-model reconstruction of QVAC's judge, not their exact serving protocol.
• +4.3% package vs the flagship imat build is the price of the Q8_0 embedding; DomCal-Slim holds the small-size end.

Links

This model is one of four verified VisionPsy-Nano GGUF packages released together under the simoneschiavoi Hugging Face namespace.

• Project website (full interactive research write-up): https://simoneschiavoi.github.io/visionpsy-optimization/
• Benchmarks (full same-harness ledger, all models × 17 benchmarks): https://simoneschiavoi.github.io/visionpsy-optimization/#benchmarks
• Hugging Face namespace (all four packages): https://huggingface.co/simoneschiavoi
• Sibling models:
  • VisionPsy-Nano DomCal
  • VisionPsy-Nano DomCal-Slim
  • VisionPsy-Nano TriStack

License and attribution

Apache-2.0 derivative. The Apache-2.0 NOTICE distributed with the artifact must be retained, and QVAC's VisionPsy-Nano-460M must be attributed as the base model. Build evidence, evaluation ledger, and reproduction scripts: https://github.com/simoneschiavoi/visionpsy-optimization. All four release packages: https://huggingface.co/simoneschiavoi.