Sign in to view source links and access this dataset
Description
REAP observer output captures per-token routing decisions and expert activation norms for every MoE layer in the moonshotai/Kimi-K2.6 model. The dataset, authored by 0xSero, contains the results of a full calibration pass, providing saliency ingredients for analysis. It was last updated on April 23, -2026.
Use Cases
Analyze expert load balancing and token routing patterns based on per-token routing decisions.
Study the relationship between input saliency and expert activation based on REAP saliency ingredients.
Benchmark layerwise activation statistics for model interpretability research based on expert activation norms.
Investigate calibration methodologies for MoE models based on the full REAP calibration pass data.
Strengths
Provides internal model statistics from a specific, named base model (moonshotai/Kimi-K2.6).
Captures data for every mixture-of-experts (MoE) layer in the model.
Includes multiple observation types: routing decisions, activation norms, and saliency ingredients.
Limitations
Column-level documentation is absent; field semantics must be inferred after download.
Row count is unknown, which may limit suitability assessment.
Description metadata is limited; actual data quality requires manual inspection after download.
Provenance
Source
moonshotai/Kimi-K2.6 model, observed via the CerebrasResearch/reap layerwise observer.
Collection Method
Output from a full REAP calibration pass (PR #17).
Time Range
null
Freshness
Last updated 2026-04 23 16:10:47; freshness should be verified.
Geography
null
License is unknown; restrictions should be verified before use.