Modelpublic

ManniX-ITA/Qwen3.5-4B-M5-OMv2-LRP

sourceHugging Faceapache-2.0updated 5mo agoView on Hugging Face

1likes132downloads

Model Card

Qwen3.5-4B-M5-OMv2-LRP

OMv2 recipe (OBIM-lite + DAREx-q + EMR election) with AttnLRP relevance scores as the importance signal driving DAREx-q sparsification. Best MBPP result of the study (51.40%) and best balanced score (53.05% / 51.40%).

Apples-to-apples comparison against M3 (OMv2 + Fisher) — same merger, same recipe, only the importance source differs (LRP relevance vs Fisher squared-grad). And against M4 (mergekit's ex-LRP) — same LRP signal, different merger.

Sources

Role	Model
Base	`Qwen/Qwen3.5-4B`
Source A	`Jackrong/Qwen3.5-4B-Claude-4.6-Opus-Reasoning-Distilled-v2`
Source B	`Crownelius/Crow-4B-Opus-4.6-Distill-Heretic_Qwen3.5`

Weights: 0.55 (A) / 0.45 (B). Density: 0.53. Seed: 42.

Phase 1+2 comparison (Q6_K)

#	Recipe	Merger	Importance	HumanEval pass@1	MBPP pass@1
floor	Qwen/Qwen3.5-4B (no merge)	—	—	60.37% 🥇	45.80%
src A	Jackrong-v2 (Claude-4.6-Opus reasoning distill)	—	—	60.37%	45.00%
src B	Crow-4B (Opus-4.6 distill, heretic)	—	—	54.88%	48.20%
M1	Vanilla DARE-TIES	daretiesmerge.py	none	51.22%	47.00%
M2	OMv2 recipe (OBIM-lite + DAREx-q + EMR election)	daretiesmerge.py	none	52.44%	49.40%
M3	OMv2 + Fisher	daretiesmerge.py	Fisher	57.93%	48.80%
M4	ex-LRP (mergekit PR #682)	mergekit (PR #682)	LRP	51.22%	49.40%
M4-v2	ex-LRP (PR #682 turbo, w=1/1, d=0.7)	mergekit PR #682 turbo	LRP	55.49%	52.20% 🥇
M5 (this)	OMv2 + LRP	daretiesmerge.py	LRP	53.05%	51.40%

Eval methodology: llama-server (--reasoning-format deepseek --reasoning-budget 8192 --parallel 2 --cache-type-k q8_0 --cache-type-v q8_0 -c 32768) → lm_eval local-completions against raw /v1/completions, temperature 0, maxgentoks=2048. All five variants scored under identical conditions. Floor reference: the top three rows are the base model and the two source fine-tunes evaluated under the same methodology. Hard finding: no merge beats `Qwen3.5-4B` base on HumanEval (best merge M3 = 57.93%, base = 60.37%). On MBPP, M4-v2 (52.20%) and M5 (51.40%) clearly beat both sources (best source: Crow at 48.20%) — merging does add MBPP capability, but borrows HumanEval ground.

LRP signal

Under lrp/ you'll find the AttnLRP relevance score safetensors used for this merge (bare-keyed form: model.X, what dare_ties_merge.py expects when the model is loaded as Qwen3_5ForConditionalGeneration and named_parameters() yields the inner-LM short paths). One file per source.

The same raw scores in multimodal-prefixed form (model.language_model.X) are published under lrp/ in the M4 repo. The two are losslessly interconvertible via scripts/rename_lrp_keys_for_multimodal.py.

Other variants in this study

—Qwen3.5-4B-M1-Dare-Ties — vanilla DARE-TIES
—Qwen3.5-4B-M2-OMv2 — OMv2 recipe (no importance signal)
—Qwen3.5-4B-M3-Fisher — OMv2 + Fisher
—Qwen3.5-4B-M4-ex-LRP — mergekit PR #682 ex-LRP
—Qwen3.5-4B-M4-v2-ex-LRP-turbo — mergekit PR #682 turbo retest
—Qwen3.5-4B-M5-OMv2-LRP — OMv2 + LRP