buugiiiiii

GLM-4.7-Flash-KGuard-REAP-32E

What this checkpoint shows

Method: K-Guard REAP
Korean MC macro: 51.2 (base 68.8, -17.5 pp)
Academic 10-benchmark macro: 36.9 (base 65.5, -28.6 pp)
Coding 2-benchmark macro: 63.4 (base 76.0, -12.6 pp)
Protection budget: 20 layers exceeded the 32-Expert budget; 185 protected Experts were force-dropped.

Benchmark results

Table with columns: benchmark, base, K-Guard 12.5%, K-Guard 25%, K-Guard 50%, REAP-Code 25%
benchmark	base	K-Guard 12.5%	K-Guard 25%	K-Guard 50%	REAP-Code 25%
Korean (held-out harness)
KMMLU	49.4	48.8	42.4	23.6	35.0
KoBEST BoolQ	90.2	87.7	78.8	74.3	79.6
KoBEST COPA	70.7	69.7	65.0	55.9	51.7
KoBEST SentiNeg	74.8	71.8	70.5	53.7	49.6
KoBEST HellaSwag	58.8	55.0	57.4	48.8	51.6
English reasoning / knowledge
ARC-Challenge	87.1	86.1	79.9	36.9	79.1
HellaSwag (en)	74.2	67.4	55.3	32.6	61.7
WinoGrande	67.4	65.6	59.2	53.1	63.5
MMLU-Pro	41.2	36.9	34.2	14.8	33.8
TruthfulQA MC2	47.3	46.0	43.1	40.6	44.6
MMLU-en	69.7	64.5	56.6	30.1	59.6
Math
GSM8K	86.3	86.7	81.8	61.3	86.7
MathQA	30.5	29.1	24.8	21.3	28.7
Coding
HumanEval+	81.1	75.6	74.4	65.2	80.5
MBPP+	70.9	69.3	67.5	61.6	67.5
Safety / bias
BBQ	76.6	69.9	59.8	33.6	74.0
SafetyBench	74.8	73.0	69.1	44.3	72.1
Macro
Korean MC macro	68.8	66.6	62.8	51.2	53.5
Academic 10-bench macro	65.5	62.5	56.4	36.9	60.4
Coding 2-bench macro	76.0	72.5	70.9	63.4	74.0

Compatibility

Runtime: Hugging Face Transformers 5.0.0 or newer, PyTorch/CUDA; validated on DGX Spark.
Architecture: Glm4MoeLiteForCausalLM.
Model code: native Transformers implementation; no repository-side modeling_*.py is required, and the checkpoint loads with trust_remote_code=False.
Preserved: first Dense layer, shared Experts, topk_method=noaux_tc, norm_topk_prob=true, router correction bias, and routed_scaling_factor=1.8.
Routing: top-4 remains unchanged; only routed Expert tensors and corresponding router rows/bias entries were compacted with one old→new ID map per layer.
The extra layer-47 MTP/NextN tensors are retained for checkpoint completeness, but runtime validation reports mtp_compatible=false. Use standard autoregressive generation; speculative MTP/NextN decoding is not supported by this release.

How it was made

Source checkpoint

Base: zai-org/GLM-4.7-Flash
Frozen revision: 7dd20894a642a0aa287e9827cb1a1f7f91386b67
Original routed MoE layout: first layer Dense, then 46 routed MoE layers with 64 Experts, top-4 routing, plus shared Experts.

Calibration dataset

Table with columns: Group, Records, Role
Group	Records	Role
Korean/English semantic pairs	2,752	Korean↔English routing contrast
Common Korean	2,752	Korean routing coverage
General replay	1,344	General-capability retention
GLM-specialized	1,344	Documents, language retention, reasoning, security, structured output, tool calling
Total	8,192	5,472 Korean / 2,720 English; 1,376 pairs

6,848 records contain source teacher responses.
1,344 GLM-specialized records use deterministic base-model rollouts.
Prompt/response observations cover 3,898,756 token memberships.
The manifest passed independent token-length and exact-prompt contamination checks.

Observation and scoring

Selected-only observer: only routed Expert assignments were measured; no dense [experts, tokens, hidden] activation tensor was retained.
REAP contribution: normalized dispatch weight × L2 norm of the unweighted Expert output.
Router selection and dispatch were checked separately so GLM's correction bias affects selection without being mistaken for the actual dispatch weight.
The observed model remained logit-identical to the uninstrumented eager model in validation.

Pruning

Uniform routed Experts per layer: 64 → 32.
Shared Experts and the first Dense layer were not pruned.
Top-4 and routed scaling factor 1.8 were retained.
Safetensors surgery physically compacted Expert tensors, router rows, and correction-bias entries.
Protection result: 20 layers exceeded the 32-Expert budget; 185 protected Experts were force-dropped.

Evaluation scope

These are decision-grade experimental results, not release-certified leaderboard numbers:

KMMLU: deterministic 500-example subset (seed 42).
KoBEST: declared BoolQ, COPA, SentiNeg, and HellaSwag evaluation splits.
ARC-Challenge, HellaSwag-en, WinoGrande, MMLU-Pro, TruthfulQA MC2, MMLU-en, GSM8K, MathQA, BBQ, and SafetyBench: deterministic 512-example subsets.
HumanEval+: full 164 tasks; MBPP+: frozen full 378-task revision; greedy pass@1.
LiveCodeBench was excluded from this run.
decision_eligible=true, release_eligible=false; do not compare these subset scores directly with full-set or few-shot leaderboard values.

Machine-readable results and the exact survivor manifest are included under evaluation/.

Certification status: 50% pruning is outside the plan's initial GLM release grid. It force-dropped 185 protected Experts and is published only as a compression-ceiling artifact.

Usage

Install a Transformers release that includes Glm4MoeLiteForCausalLM:

bash
pip install "transformers>=5.0.0" accelerate safetensors

The checkpoint was validated with Transformers 5.0.0 and NVIDIA PyTorch 25.09. It uses the native Transformers model implementation and does not execute custom code from this repository.

python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "buugiiiiii/GLM-4.7-Flash-KGuard-REAP-32E"
tokenizer = AutoTokenizer.from_pretrained(
    model_id,
    trust_remote_code=False,
)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=False,
    attn_implementation="eager",
    experts_implementation="eager",
)
model.eval()

messages = [{"role": "user", "content": "한국어로 이진 탐색을 설명해줘."}]
prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=False,
)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.inference_mode():
    output = model.generate(
        **inputs,
        max_new_tokens=256,
        do_sample=False,
        pad_token_id=tokenizer.eos_token_id,
    )

print(tokenizer.decode(output[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Limitations

Expert pruning is irreversible within this checkpoint and can change routing behavior sharply on tasks not represented by the calibration or evaluation sets. No post-pruning recovery or fine-tuning was applied. Reported Korean behavior is relative to the tested base checkpoint and harness; it does not imply universal Korean-language safety or quality preservation. The current release is not validated for speculative MTP/NextN decoding.

Included research artifacts

evaluation/glm-full-b19-results.json: machine-readable five-model benchmark results
evaluation/glm-full-b19-long-scores.csv: long-format benchmark scores
evaluation/survivor-manifest.json: exact per-layer old Expert IDs retained
artifacts/calibration-validation.json: independent calibration-manifest validation
artifacts/observation-validation.json: selected-only observation validation
artifacts/expert-statistics.csv: per-Expert scoring/features used for this method
: exact K-Guard protected set and evidence

Provenance

Base model: zai-org/GLM-4.7-Flash, revision 7dd20894a642a0aa287e9827cb1a1f7f91386b67
Base license: MIT
Research code: K-REAP
Project: K-Guard-REAP
Team: RAS-LAB
Benchmark report SHA256: 7f7ac2f767e68c61528082e62c4eef3a6027429ab1e47b654de10b4c93b760d6
Model card generated from a benchmark report with status PASS and zero validation failures.

Citation

Please cite the original GLM work and the K-Guard/K-REAP repository when using this checkpoint.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

buugiiiiii

Model Tree

Base

zai-org/GLM-4.7-Flash

Fine-tuned

this model

Input Modalities

Text

Output Modalities

Text

Supported Functionality

Dedicated EndpointsContainer

Explore FriendliAI today

Get started Talk to an engineer

What this checkpoint shows

Method: K-Guard REAP
Korean MC macro: 51.2 (base 68.8, -17.5 pp)
Academic 10-benchmark macro: 36.9 (base 65.5, -28.6 pp)
Coding 2-benchmark macro: 63.4 (base 76.0, -12.6 pp)
Protection budget: 20 layers exceeded the 32-Expert budget; 185 protected Experts were force-dropped.

Benchmark results

Table with columns: benchmark, base, K-Guard 12.5%, K-Guard 25%, K-Guard 50%, REAP-Code 25%
benchmark	base	K-Guard 12.5%	K-Guard 25%	K-Guard 50%	REAP-Code 25%
Korean (held-out harness)
KMMLU	49.4	48.8	42.4	23.6	35.0
KoBEST BoolQ	90.2	87.7	78.8	74.3	79.6
KoBEST COPA	70.7	69.7	65.0	55.9	51.7
KoBEST SentiNeg	74.8	71.8	70.5	53.7	49.6
KoBEST HellaSwag	58.8	55.0	57.4	48.8	51.6
English reasoning / knowledge
ARC-Challenge	87.1	86.1	79.9	36.9	79.1
HellaSwag (en)	74.2	67.4	55.3	32.6	61.7
WinoGrande	67.4	65.6	59.2	53.1	63.5
MMLU-Pro	41.2	36.9	34.2	14.8	33.8
TruthfulQA MC2	47.3	46.0	43.1	40.6	44.6
MMLU-en	69.7	64.5	56.6	30.1	59.6
Math
GSM8K	86.3	86.7	81.8	61.3	86.7
MathQA	30.5	29.1	24.8	21.3	28.7
Coding
HumanEval+	81.1	75.6	74.4	65.2	80.5
MBPP+	70.9	69.3	67.5	61.6	67.5
Safety / bias
BBQ	76.6	69.9	59.8	33.6	74.0
SafetyBench	74.8	73.0	69.1	44.3	72.1
Macro
Korean MC macro	68.8	66.6	62.8	51.2	53.5
Academic 10-bench macro	65.5	62.5	56.4	36.9	60.4
Coding 2-bench macro	76.0	72.5	70.9	63.4	74.0

Compatibility

Runtime: Hugging Face Transformers 5.0.0 or newer, PyTorch/CUDA; validated on DGX Spark.
Architecture: Glm4MoeLiteForCausalLM.
Model code: native Transformers implementation; no repository-side modeling_*.py is required, and the checkpoint loads with trust_remote_code=False.
Preserved: first Dense layer, shared Experts, topk_method=noaux_tc, norm_topk_prob=true, router correction bias, and routed_scaling_factor=1.8.
Routing: top-4 remains unchanged; only routed Expert tensors and corresponding router rows/bias entries were compacted with one old→new ID map per layer.
The extra layer-47 MTP/NextN tensors are retained for checkpoint completeness, but runtime validation reports mtp_compatible=false. Use standard autoregressive generation; speculative MTP/NextN decoding is not supported by this release.

How it was made

Source checkpoint

Base: zai-org/GLM-4.7-Flash
Frozen revision: 7dd20894a642a0aa287e9827cb1a1f7f91386b67
Original routed MoE layout: first layer Dense, then 46 routed MoE layers with 64 Experts, top-4 routing, plus shared Experts.

Calibration dataset

Table with columns: Group, Records, Role
Group	Records	Role
Korean/English semantic pairs	2,752	Korean↔English routing contrast
Common Korean	2,752	Korean routing coverage
General replay	1,344	General-capability retention
GLM-specialized	1,344	Documents, language retention, reasoning, security, structured output, tool calling
Total	8,192	5,472 Korean / 2,720 English; 1,376 pairs

6,848 records contain source teacher responses.
1,344 GLM-specialized records use deterministic base-model rollouts.
Prompt/response observations cover 3,898,756 token memberships.
The manifest passed independent token-length and exact-prompt contamination checks.

Observation and scoring

Selected-only observer: only routed Expert assignments were measured; no dense [experts, tokens, hidden] activation tensor was retained.
REAP contribution: normalized dispatch weight × L2 norm of the unweighted Expert output.
Router selection and dispatch were checked separately so GLM's correction bias affects selection without being mistaken for the actual dispatch weight.
The observed model remained logit-identical to the uninstrumented eager model in validation.

Pruning

Uniform routed Experts per layer: 64 → 32.
Shared Experts and the first Dense layer were not pruned.
Top-4 and routed scaling factor 1.8 were retained.
Safetensors surgery physically compacted Expert tensors, router rows, and correction-bias entries.
Protection result: 20 layers exceeded the 32-Expert budget; 185 protected Experts were force-dropped.

Evaluation scope

These are decision-grade experimental results, not release-certified leaderboard numbers:

KMMLU: deterministic 500-example subset (seed 42).
KoBEST: declared BoolQ, COPA, SentiNeg, and HellaSwag evaluation splits.
ARC-Challenge, HellaSwag-en, WinoGrande, MMLU-Pro, TruthfulQA MC2, MMLU-en, GSM8K, MathQA, BBQ, and SafetyBench: deterministic 512-example subsets.
HumanEval+: full 164 tasks; MBPP+: frozen full 378-task revision; greedy pass@1.
LiveCodeBench was excluded from this run.
decision_eligible=true, release_eligible=false; do not compare these subset scores directly with full-set or few-shot leaderboard values.

Machine-readable results and the exact survivor manifest are included under evaluation/.

Certification status: 50% pruning is outside the plan's initial GLM release grid. It force-dropped 185 protected Experts and is published only as a compression-ceiling artifact.

Usage

Install a Transformers release that includes Glm4MoeLiteForCausalLM:

bash
pip install "transformers>=5.0.0" accelerate safetensors

The checkpoint was validated with Transformers 5.0.0 and NVIDIA PyTorch 25.09. It uses the native Transformers model implementation and does not execute custom code from this repository.

python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "buugiiiiii/GLM-4.7-Flash-KGuard-REAP-32E"
tokenizer = AutoTokenizer.from_pretrained(
    model_id,
    trust_remote_code=False,
)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=False,
    attn_implementation="eager",
    experts_implementation="eager",
)
model.eval()

messages = [{"role": "user", "content": "한국어로 이진 탐색을 설명해줘."}]
prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=False,
)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.inference_mode():
    output = model.generate(
        **inputs,
        max_new_tokens=256,
        do_sample=False,
        pad_token_id=tokenizer.eos_token_id,
    )

print(tokenizer.decode(output[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Limitations

Included research artifacts

evaluation/glm-full-b19-results.json: machine-readable five-model benchmark results
evaluation/glm-full-b19-long-scores.csv: long-format benchmark scores
evaluation/survivor-manifest.json: exact per-layer old Expert IDs retained
artifacts/calibration-validation.json: independent calibration-manifest validation
artifacts/observation-validation.json: selected-only observation validation
artifacts/expert-statistics.csv: per-Expert scoring/features used for this method
: exact K-Guard protected set and evidence

Provenance

Base model: zai-org/GLM-4.7-Flash, revision 7dd20894a642a0aa287e9827cb1a1f7f91386b67
Base license: MIT
Research code: K-REAP
Project: K-Guard-REAP
Team: RAS-LAB
Benchmark report SHA256: 7f7ac2f767e68c61528082e62c4eef3a6027429ab1e47b654de10b4c93b760d6
Model card generated from a benchmark report with status PASS and zero validation failures.

Citation

Please cite the original GLM work and the K-Guard/K-REAP repository when using this checkpoint.

GLM-4.7-Flash-KGuard-REAP-32E

README

What this checkpoint shows

Benchmark results

Compatibility

How it was made

Source checkpoint

Calibration dataset

Observation and scoring

Pruning

Evaluation scope

Usage

Limitations

Included research artifacts

Provenance

Citation

Explore FriendliAI today

README

What this checkpoint shows

Benchmark results

Compatibility

How it was made

Source checkpoint

Calibration dataset

Observation and scoring

Pruning

Evaluation scope

Usage

Limitations

Included research artifacts

Provenance

Citation