| slices: | |
| - sources: | |
| - model: Theros/Qwen2.5-ColdBrew-R1-test3 | |
| layer_range: [0, 28] | |
| - model: bunnycore/Qwen-2.5-7B-Stock-Deep-Bespoke-v2 | |
| layer_range: [0, 28] | |
| merge_method: slerp | |
| base_model: Theros/Qwen2.5-ColdBrew-R1-test3 | |
| parameters: | |
| t: | |
| - filter: self_attn | |
| value: [0.3, 0.5, 0.6, 0.6, 0.7] # Avoids extreme low/high fluctuations | |
| - filter: mlp | |
| value: [0.7, 0.6, 0.5, 0.4, 0.3] # Gradual shift, avoiding an early MLP spike | |
| - value: 0.5 | |
| dtype: bfloat16 | |
| tokenizer_source: union |