Darwin-V9-Chimera-4B (Generation 2)

VIDRAFT attention + Qwen3-4B / Gemma4-E4B FFN crossbreed. A Qwen3-4B × Gemma4-E4B hybrid — NOT from-scratch. Private research checkpoint.

Lineage (Darwin-Chimera 계보)

Gen	Model	Composition
Gen 1	Darwin-Chimera-4B-Gen1	Qwen3-4B attention-healing adapter (FFN = Qwen3-4B, frozen)
Gen 2 (this)	Darwin-V9-Chimera-4B	Gen1 adapter + Gemma4-E4B FFN crossbreed → re-healing

What this is

The Gen-1 adapter's FFN is reconstructed by cross-breeding Qwen3-4B FFN with Gemma4-E4B FFN (ratio 0.15), then the attention is re-healed (VIDRAFT) to adapt to the fused FFN. This carries the Gen-1 attention forward while blending a second model's knowledge — so the result is not reducible to any single parent.

attention: VIDRAFT healing (Qwen3-4B based)
FFN: Qwen3-4B 85% ⊕ Gemma4-E4B 15% (bilinear inter projection 10240→9728, layer map 42→36)
structure: 2560 / 9728 / 36L (Qwen3-4B coordinates)
re-healing: 0.5B tokens, attention-only, LR 1e-5

Evaluation (same harness, base zero-shot, KMMLU 3 subjects / 90Q)

model	KMMLU	stage
Qwen3-4B (original)	13.3%	base
Gemma4-E4B (base)	26.7%	base
Darwin-Chimera Gen1	27.1%*	base
fused raw (pre re-heal)	22.2%	intermediate
Darwin-V9 (this)	27.8%	base

* Gen1 measured on 6 subjects. All numbers are base zero-shot — instruction-following quality is expected from a later SFT stage (cf. Gemma4-E4B base 26.7% → it 69.4%).

→ After blending 15% Gemma4 FFN, performance is maintained / slightly above the Gen-1 baseline and Gemma4-E4B base. Gemma knowledge is visibly incorporated (multilingual facts, "Germany is Berlin / Italy is ..."), and the intermediate English degradation is recovered by re-healing.

Known limitations

Some Korean repetition remains in greedy single-shot generation → to be resolved by SFT.
Absolute scores are base-level; this is a research backbone, not a production/instruct model.

License

Gemma Terms of Use (Gemma4-E4B weights are blended in) + Apache 2.0 (Qwen3-4B). Built on Qwen/Qwen3-4B and google/gemma-4-E4B.