Darwin V6 Evolved Model

Created by Darwin V6 diagnostic-guided evolutionary merge engine.

Parent Models

• Father: Qwen/Qwen3.5-27B
• Mother: Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled

Evolution Result

• Benchmark score: 0.8162
• Merge method: slerp
• Merge hash: 0ae8b8cb

Merge Statistics

• Total tensors merged: 1199
• Transplant A (Father preserved): 0
• Transplant B (Mother preserved): 0
• Blended: 1199

Optimal Genome

 global_ratio: 0.4893
 attn_ratio: 0.1463
 ffn_ratio: 0.8768
 embed_ratio: 0.3021
 density_a: 0.8507
 density_b: 0.9413
 block_0_ratio: 0.6750
 block_1_ratio: 0.4790
 block_2_ratio: 0.3610
 block_3_ratio: 0.4365
 block_4_ratio: 0.5255
 block_5_ratio: 0.6525
 block_6_ratio: 0.5557
 block_7_ratio: 0.2783
 block_8_ratio: 0.7302
 block_9_ratio: 0.4764
 mri_trust: 0.8571
 merge_method_weight: 0.7802

MRI Prescription Summary

• Average ratio_b: 0.500
• Attention ratio: 0.500
• FFN ratio: 0.500
• Embed ratio: 0.500
• Transplant A: 0
• Transplant B: 0
• Blended: 1199

Health Check

failed: Repo id must be in the form 'repo_name' or 'namespace/repo_name': '/data/ray_temp/ginipick/darwin_merge_cache/merged_9cd05ed7'. Use repo_type argument if needed.

Method

Darwin V6 implements DARE-TIES merge directly via PyTorch tensor operations. Per-tensor ratios are determined by MRI diagnostic (static tensor analysis + probe-based functional importance) combined with evolutionary genome search.

Formula: final_ratio = mri_ratio * mri_trust + genome_ratio * (1 - mri_trust)

DARE-TIES algorithm: Yadav et al., 2023 (re-implemented, not library-dependent)

Built by VIDRAFT. Apache 2.0.

Base model

Qwen/Qwen3.5-27B

Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled

(38)