simoneschiavoi/VisionPsy-Nano-DomCal

VisionPsy-Nano DomCal

The q4_0 envelope, repaired at zero measured cost: 12W/4L/1T over 17 descriptive rows vs same-harness QVAC q4_0 — pooled +2.2 points (p = 4.7×10⁻⁵) — and 2W/15L/0T over 17 displayed rows vs QVAC q4_k_m-imat (same-harness re-run) at −44.6 MiB (−11.4%).

A GGUF quantized derivative of QVAC's VisionPsy-Nano-460M built with P1 (domcal) — domain-calibrated imatrix quantization. This is not a new trained checkpoint: the ~460M-parameter architecture (SigLIP2 vision encoder + SmolLM2-360M backbone) is unchanged; only the quantization calibration was modified. Part of the VisionPsy-Nano release collection — see the Links section below.

Exact runtime pair

Component	File	Bytes
LM	visionpsy-nano-460m-q4_0-domcal.gguf	256379776
mmproj	mmproj-visionpsy-nano-460m-q8.gguf	108782144

Package size: 348.2 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)
Technique	P1 (domcal) — importance matrix computed on a domain-matched corpus, steering q4_0 rounding toward the tensors that matter for visual question answering
Quantization	q4_0 LM (same format as QVAC's baseline) + QVAC's own q8 mmproj — same size class as the base
Calibration data	VQAv2-train + TextVQA-train, ChatML-formatted; training splits only — never benchmark test data, no evaluation images
Total size	348.2 MiB = 244.5 MiB LM + 103.7 MiB mmproj — +0.5 MiB vs QVAC q4_0 (347.7), −44.6 MiB (−11.4%) vs QVAC q4_k_m-imat (392.8)
Headline	12W/4L/1T over 17 descriptive rows vs same-harness QVAC q4_0; 9W/4L/1T over 14 non-selection rows; normalized 60.31 vs q4_0's 59.48; 2W/15L/0T over 17 displayed rows vs QVAC q4_k_m-imat (same-harness re-run)
Observed speed	802.2 ms/item vs q4_0's 803.5 — repair at zero measured latency cost (−1.3 ms/item)

Why this build exists

The study's core finding: a label like "4-bit" says how many bits are available, not whether those bits protect the tensors that matter for the model's actual job. Domain calibration is a targeted repair, not a universal upgrade. Applied to QVAC's uncalibrated q4_0 build, it recovered a pooled +2.2 points (p = 4.7×10⁻⁵) — the largest single-technique gain measured in the campaign. The same idea applied to QVAC's already-calibrated q4_k_m did not stack, and at 5-bit it actively hurt (strict OCR −2.2, p = 0.003): calibration matters most where importance information is missing. DomCal is that finding, shipped: the repair applied exactly where it helps.

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	2W/15L/0T	12W/4L/1T (17 descriptive rows)
Package size Δ (348.2 MiB)	−44.6 MiB (−11.4%)	+0.5 MiB
Speed Δ (802.2 ms/item)	−11.0 ms/item (−1.4%)	−1.3 ms/item

Measured results

Strongest gains land exactly where the uncalibrated build was weakest: MMStar 45.8 vs 42.3, MMBench 58.1 vs 55.0, MathVista 43.9 vs 41.7 against the q4_0 control. The named losses stay visible: QVAC q4_0 keeps MME, POPE and RealWorldQA; MMMU is tied. Against the flagship imat build, DomCal posts an honest 2W/15L/0T — imat keeps OCRBench, DocVQA, ChartQA and InfoVQA. This is the same-format repair of the q4_0 baseline and a smaller, faster alternative to the flagship — not an imat-beater.

Limitations

• Exploratory scope: one seed (17), one harness, one GPU. Counts are descriptive rows, not statistical proof or universal-superiority claims; the pooled +2.2-point repair against q4_0 is the pre-registered significance claim (p = 4.7×10⁻⁵).
• Judged rows: qwen3.6-27b via OpenRouter — an attempted same-model reconstruction of QVAC's judge, not their exact serving protocol.
• Calibration scope: benefits demonstrated on the uncalibrated q4_0 base; the technique is not additive on already-calibrated builds.
• A completed post-hoc contamination audit of the calibration corpus is summarized in the release ledger (release repository).

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-Slim
  • VisionPsy-Nano DomCal-EmbGuard
  • 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.