qwen3-4b-vlbackbone-text

The language backbone of Qwen/Qwen3-VL-4B-Instruct, extracted as a standalone text model.

The vision tower (model.visual.*) and multimodal projector are removed; the text decoder (model.language_model.*) + lm_head are repackaged as a stock Qwen3ForCausalLM. The motivation is a substrate choice for the Chladni project: keep the visuospatially-shaped representations the text decoder learned while staying consistent with a visual stream, but train pure text through the standard (non-vision) fine-tuning path.

Faithfulness

Verified bit-exact against the original VL model's text-only forward: max |Δ logit| = 0.00e+00 and 100% next-token argmax agreement across text, code, interior-monologue, and ASCII-grid (ARC) probes. The one subtlety — Qwen3-VL stores rope_theta inside rope_scaling — is handled so the standard 1-D RoPE reproduces the model's interleaved MRoPE for text-only sequences.

Produced by tools/extract_vl_text_backbone.py (Kenneth Grace & Kira).