| name: Q2.5-ColdBrew-R1-Indigo | |
| const_tag: &scale_factor 0.7071067812 # 1/sqrt(2) scaling for stability | |
| attenuate-env: | |
| parameters: | |
| scale: | |
| - filter: q_proj | |
| value: | |
| - filter: k_proj | |
| value: | |
| - value: 1.0 | |
| slices: | |
| - sources: | |
| - model: Theros/Qwen2.5-ColdBrew-R1 | |
| layer_range: [0, 8] # Retaining foundational knowledge and language structure. | |
| - sources: | |
| - model: Theros/Qwen2.5-ColdBrew-R1 | |
| layer_range: [9, 19] # Full-strength duplication of mid-range reasoning layers. | |
| - sources: | |
| - model: Theros/Qwen2.5-ColdBrew-R1 | |
| layer_range: [10, 19] # Targeted reinforcement, slightly attenuated to avoid over-dominance. | |
| <<: | |
| - sources: | |
| - model: Theros/Qwen2.5-ColdBrew-R1 | |
| layer_range: [20, 28] # Keeping higher-level abstract processing untouched for stability. | |
| merge_method: passthrough | |
| dtype: bfloat16 | |
| normalize: true | |
| int8_mask: true | |