CoolFace
Datasetpublic

philipjohnbasile/glm52-demolition-data

GLM-5.2-Demolition — Training & Calibration Data Apple Silicon AI hub · Model release · MLX code sample Preview scope, checked September 10, 2026: the default Hub viewer indexes 87,586 rows (84,231 train, 3,277 validation, 78 test). The original release total below describes the broader JSONL repository. Use the file browser and explicit file selections when reusing a particular corpus. The hub includes a checked download example for the seven-row MLX code sample. The data… See the full description on the dataset page: https://huggingface.co/datasets/philipjohnbasile/glm52-demolition-data.

sourceHugging Facemitupdated 12d agoView on Hugging Face
3likes288downloads
Dataset Card

GLM-5.2-Demolition — Training & Calibration Data

Apple Silicon AI hub · Model release · MLX code sample

Preview scope, checked September 10, 2026: the default Hub viewer indexes 87,586 rows (84,231 train, 3,277 validation, 78 test). The original release total below describes the broader JSONL repository. Use the file browser and explicit file selections when reusing a particular corpus. The hub includes a checked download example for the seven-row MLX code sample.

The data behind `philipjohnbasile/GLM-5.2-Demolition-q4a4-soul-MLX`. 272,549 examples (~1.05 GB) across three uses: LoRA-heal gold, swappable-soul gold, and prune calibration.

Why this dataset matters more than the weights. A measured finding of the project (see the model's MISSION_SUMMARY.md): the data re-derives the adapters; the adapters re-derive from nothing. The weights are a disposable artifact of one base — this data is the reusable asset, and it carries over to heal any clean base (e.g. Qwen3-Coder-30B).

Contents

GroupWhatUse
heal/ gold (vanilla FOCUS-9)verified code in Python · TypeScript · JavaScript · Rust · Go · HTML · CSS · SQL · Postgres — vanilla, latest versions, no frameworksLoRA-heal the pruned base back to coherence
heal/souls_\*heritage "soul" gold per facet — security, science, legacy, design, art, music, perfumery, fullstack, gamedevper-facet swappable LoRA adapters (the "model factory")
calib/calibration corpus (incl. reap_mix.jsonl)REAP saliency — which experts to keep when pruning
  • —Format: JSONL, OpenAI chat schema ({"messages":[{"role","content"},...]}).
  • —Verification: the code gold is flywheel-generated and verifier-checked (compiled/tested via the pipeline's src/verifiers.py) — not scraped, not unchecked model output.
  • —FOCUS-9 + souls split: the core is pure vanilla FOCUS-9; everything applied (frameworks, specialties) lives in the swappable soul gold, never the core.

Provenance & honesty

  • —Distilled/flywheeled from the MIT-licensed `zai-org/GLM-5.2` pipeline, filtered to vanilla FOCUS-9 + verified.
  • —Known gaps logged honestly: Go/JS vanilla gold is thinner than Python; the security-soul heal measured neutral-to-degrading on the demolished base (it's gold for a clean base, not a proven lift on the artifact).

How to use

python
from datasets import load_dataset
ds = load_dataset("philipjohnbasile/glm52-demolition-data", data_files="heal/**/*.jsonl")

Heal a clean base with the FOCUS-9 gold, mount a soul from heal/souls_*, or reuse calib/ to prune your own MoE.

License

MIT (GLM-5.2 is Z.ai "Pure Open"). Attribute Z.ai + this repo if you build on it.