Self-Abliterated Qwen3-VL-4B-Instruct — Krea2 Text-Encoder Replacement

A refusal-removed ("abliterated") copy of Qwen/Qwen3-VL-4B-Instruct, produced with Heretic.

Why this exists

Krea2 (an image-generation model) uses Qwen3-VL-4B as its frozen text/CLIP encoder — Krea2's own prompt adherence depends entirely on that encoder's internal representations, across both its TextGenerate prompt-refinement path and its CLIPTextEncode embedding-conditioning path. Alignment training suppresses how restricted content is represented internally, not just how it's generated — this checkpoint removes that suppression via the same single-pass Heretic technique already validated for this project's own production Qwen3.5-9B base (ADR-004).

What this is (and isn't)

This is a drop-in replacement for Krea2's own qwen3vl_4b_fp8_scaled.safetensors CLIP loader input — not a video-captioning base for finetune-vlm/--caption-model auto. It is a completely separate artifact from this project's own existing production Qwen3.5-9B base repo, produced for a different consumer (Krea2's own ComfyUI workflow, not this project's own caption-training pipeline) and published to its own, independent HF Hub repo.

Packaging

• Root: the standard-precision (BF16) HF-transformers-format checkpoint (Heretic's own merge export strategy).
• fp8/: the same checkpoint, cast to torch.float8_e4m3fn per-tensor for smaller VRAM footprint. This is a straightforward cast, not Krea2's own fp8_scaled format (a calibrated per-tensor scale factor computed to preserve dynamic range better than a naive cast) — disclosed honestly, not claimed as exact parity with the official qwen3vl_4b_fp8_scaled.safetensors checkpoint. A calibrated scaled-fp8 conversion, if needed, is a separate, later concern (this project's qwen3vl-encoder-abliteration epic).

License

Apache 2.0, inherited from the base model (Qwen/Qwen3-VL-4B-Instruct). Abliteration only edits existing weights (no new copyrighted material introduced), so the original license carries through.