DmitryDB/Qwen3-VL-32B-Instruct-Heretic-Adaptive-v1

Qwen3-VL-32B-Instruct Heretic Adaptive v1

This repository contains three weight-edited variants of Qwen/Qwen3-VL-32B-Instruct. They were selected from one 600-trial adaptive search. All variants preserve the original Transformers model layout and can be loaded with the same Qwen3-VL-compatible software as the base model.

Variants

Directory	Search trial	Intended trade-off
balanced	303	Recommended default; strongest measured quality preservation
max-removal	273	Lowest sparse refusal-geometry score in this search
marker-zero	262	Zero lexical refusal-marker hits on the 136-prompt search set

Trial numbers are zero-based Optuna trial IDs. Heretic's interactive export UI shows them as 304, 274, and 263 respectively.

Exact perplexity check

The final exports were reloaded and measured on the same frozen 400 x 512-token WikiText-2 stream in BF16.

Variant	Perplexity	Relative to base
Base model	8.432361	0.000%
balanced	8.426778	-0.066%
max-removal	8.428192	-0.049%
marker-zero	8.434537	+0.026%

Small negative changes are normal measurement outcomes: the edit can slightly improve likelihood on this particular finite sample. They should not be read as a general capability improvement.

Search measurements

Lower sparse refusal-geometry values are better. Lexical marker counts are an auxiliary diagnostic only: zero markers do not prove zero semantic evasion, and a positive marker count does not prove that the answer was a refusal.

Variant	Sparse refusal geometry	Positive probes	Lexical markers	Search PPL delta
max-removal	-0.009977	53 / 136	4 / 136	+0.039%
balanced	-0.008917	53 / 136	2 / 136	-0.051%
marker-zero	-0.008833	58 / 136	0 / 136	+0.335%

Reload smoke test

The base model and all three exports were independently reloaded and generated 16 responses each with the same inputs and settings. Every archive had 16 unique IDs and zero empty responses. This is a technical export/reload test, not a substitute for semantic evaluation.

Loading

Download one variant directory and load it as a normal local Qwen3-VL model.

from transformers import AutoProcessor, Qwen3VLForConditionalGeneration

path = "balanced"
model = Qwen3VLForConditionalGeneration.from_pretrained(
    path,
    torch_dtype="auto",
    device_map="auto",
)
processor = AutoProcessor.from_pretrained(path)

Use a recent Transformers build with Qwen3-VL support. No adapter is required; the published directories contain merged weights.

Limitations

• The search responses were short and the lexical metric is not a semantic judge. Soft substitutions and indirect evasions may remain.
• Perplexity covers one frozen text benchmark and does not establish unchanged vision, reasoning, multilingual, or long-context behavior.
• The three variants are alternatives, not checkpoints that should be stacked.
• Evaluate the selected variant on the prompts and generation settings relevant to your own use case.

Numeric reports and the search journal are included under evaluation/ and search/ for reproducibility. Prompt and response text is not published.