Datasetpublic

0xSero/glm5-reap50-tunecomp-sft

glm5-reap50-tunecomp-sft Supervised chat finetuning dataset used for the GLM-5 REAP-50 TuneComp LoRA recovery run. Contents training_data.jsonl: 754 chat samples manifest.json: row counts and provenance Schema Each JSONL row contains: id category messages prompt response content_length reasoning_length messages is a 3-message chat list with system, user, and assistant roles. Provenance Teacher source: remote glm-5 responses… See the full description on the dataset page: https://huggingface.co/datasets/0xSero/glm5-reap50-tunecomp-sft.

sourceHugging Faceotherupdated 25d agoView on Hugging Face

0likes31downloads

Dataset Card

glm5-reap50-tunecomp-sft

Supervised chat finetuning dataset used for the GLM-5 REAP-50 TuneComp LoRA recovery run.

—training_data.jsonl: 754 chat samples
—manifest.json: row counts and provenance

Schema

Each JSONL row contains:

—id
—category
—messages
—prompt
—response
—content_length
—reasoning_length

messages is a 3-message chat list with system, user, and assistant roles.

Provenance

—Teacher source: remote glm-5 responses collected via Z.AI API
—Transformation: converted into SFT-style chat training data
—Student base used in training: GLM-5-REAP-50pct-FP8-scalefix

Notes

This is the dataset artifact from the March 28, 2026 B200 TuneComp run.

Acknowledgements

—Observations collected from GLM-5 REAP during REAP expert-pruning runs.
—zai-org/GLM-5 for the upstream base model.
—Cerebras Research for the REAP method. TuneComp SFT data derived from the REAP lineage.