0xSero/kimi-k2.6-reap-observations-v1

该数据集包含对moonshotai/Kimi-K2.6模型进行完整REAP校准通过的观察者输出。每条记录描述了基础模型每个MoE层的每令牌路由决策、专家激活规范和REAP显著性成分。下游用户可以将这些观察结果反馈到`reap.prune`中，以在不重新运行（昂贵的）前向校准的情况下，以任意压缩比生成修剪后的检查点。源模型为Kimi-K2.6（DeepseekV3架构，约1.026 T参数），使用INT4量化，组大小为32，对称，压缩张量`pack-quantized`格式。校准使用了复合数据集和特定的REAP参数。

This dataset contains the observer output of a full REAP calibration pass on moonshotai/Kimi-K2.6. Each record describes per-token routing decisions, expert activation norms, and the REAP saliency ingredients for every MoE layer of the base model. Downstream consumers can feed these observations back into `reap.prune` to produce pruned checkpoints at arbitrary compression ratios without re-running the (expensive) forward-pass calibration. The source model is Kimi-K2.6 (DeepseekV3 arch, ~1.026 T params) with INT4 quantization, group-size 32, symmetric, compressed-tensors `pack-quantized` format. Calibration used a composite dataset and specific REAP parameters.