Files
QwenR1-7B-Breadcrumbs-TIES/mergekit_config.yml

12 lines
374 B
YAML
Raw Permalink Normal View History

merge_method: breadcrumbs_ties
base_model: deepseek-ai/DeepSeek-R1-Distill-Qwen-7B
parameters:
density: 0.15 # Retain top 15% (Trims bottom 85% SFT noise)
gamma: 0.99 # Trim the top 1% (Destroys extreme RL outliers)
weight: 1.0
models:
- model: nvidia/AceReason-Nemotron-7B
- model: ./padded-maestro
- model: POLARIS-Project/Polaris-7B-Preview
dtype: bfloat16