buugiiiiii/GLM-4.7-Flash-SameCal-REAP-56E-Control

GLM-4.7-Flash Same-Calibration REAP 56E Control

Public same-calibration standard-REAP control from the GLM-4.7-Flash K-Guard REAP study.

• —Routed experts: 64 → 56 per MoE layer (12.5% pruning)
• —Active experts per token: top-4 unchanged
• —Base revision: 7dd20894a642a0aa287e9827cb1a1f7f91386b67
• —Role: standard-REAP scoring control using the same calibration run
• —Fine-tuning: none
• —Precision: BF16
• —Status: experimental; release_eligible=false

This checkpoint exists to reduce a major confound in the main comparison: it uses the same calibration run and 56E budget as the K-Guard controls, while using standard REAP survivor scoring. It does not establish calibration-seed stability or production readiness.

B22 controlled benchmark

Values are percentages. LiveCodeBench uses v6, greedy n=1, so it is not directly interchangeable with official n=10 leaderboard scores.

The Korean macro difference between Soft K-Guard and this control was approximately 0.00 pp with a paired stratified bootstrap 95% interval of [-1.02, 1.03]. This controlled run did not support a Korean causal advantage from soft protection.

Loading

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "buugiiiiii/GLM-4.7-Flash-SameCal-REAP-56E-Control"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

Use a recent Transformers version that supports Glm4MoeLiteForCausalLM.

Limitations

• —One calibration run and one 56E budget; calibration-seed stability is unknown.
• —This is a research control, not a recommended production checkpoint.
• —Top-4 routing is unchanged, so checkpoint reduction does not by itself prove higher decode throughput.
• —The public weights support reproducibility; the experiment code and detailed manifests remain in the associated private research repository.