Skip to content

deucebucket/cerebellum

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

6 Commits
 
 
 
 
 
 
 
 

Repository files navigation

Cerebellum

Cerebellum is the public home for model releases, benchmark summaries, model cards, and high-level reproducibility notes for Cerebellum quantized GGUFs.

The public repository is intentionally scoped to safe release material:

  • model cards and release notes
  • benchmark summaries and audited result metadata
  • high-level recipes for using released GGUFs
  • public CLI surface documentation
  • release artifact hashes, runtime flags, and links

The private development repository contains the factory: pipeline automation, tensor-selection internals, raw ablation traces, experimental scripts, and working notes.

Releases — Heretic Cerebellum series (2026-06-11)

Cerebellum mixed-precision quants of community Heretic (abliterated) variants. Each repo includes the model card, full audited per-question benchmark artifacts, and adversarial audit reports. All wrong answers were individually verified as genuine model errors before publication.

Release Size Highlights
Qwen3.6-35B-A3B-Heretic-Cerebellum-GGUF 11.96 GB ARC 95.48, HellaSwag 91.78, MMLU-Redux 75.42, HumanEval+ 64.63, vision 24/24. Head-to-head: beats the 16.87 GB uniform Q3_K_M of the same weights on PPL, MMLU and HumanEval+ at 29% smaller.
Qwen3.6-27B-Heretic-Cerebellum-GGUF 12.87 GB ARC 96.93, MMLU-Redux 76.21, HumanEval+ 84.76 (chat no-think), vision 24/24, RealWorldQA 78%.
Gemma-4-E4B-it-Heretic-Cerebellum-GGUF 4.51 GB Beats stock Cerebellum v2 on ARC (87.37), MMLU-Redux (58.63) and HumanEval (70.12) at identical size.

A fourth candidate (Gemma 4 E2B heretic) failed release gates — the source abliteration measurably damaged the base model (code benchmarks -31 pts at full precision) — and was not published. Methodology note: heretic transfers are screened by the source's reported KL divergence; heavy ablations (KL >~ 0.05) require full-precision screening before quantization.

Org collections: DB-Cerebellum.

How the method works, at the public level: docs/how_cerebellum_works.md.

Benchmark Evidence

The benchmarks/ directory holds the published per-model evidence: summary JSONs and per-question detailed answer files for every benchmark number in the model cards. docs/benchmark_protocol.md documents the exact harness rules, runtime flags, and audit requirements a result must clear before it is published. Anyone with the model file and these documents can reproduce or dispute any published score. That is the point.

About

Ablation-informed mixed-precision quantization for LLMs

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors