cgg507

GLM-5.3-Flash-NVFP4-GM-step300

Behaviour changes vs. the base model

Game Master voice: second-person narration, scene beats that end by handing control back to the player.
Much shorter roleplay turns (~60% fewer characters per reply in our roleplay suite) and very brief reasoning.
Fewer soft refusals / cut-aways in dark roleplay scenes (violence, villain cruelty); commits to crude personas.
Outside roleplay, refusal behaviour on directly harmful requests appears unchanged (preliminary — see below).

Evaluation (lm-evaluation-harness 0.4.12, vLLM 0.31, 2x B200, identical settings for base and this model)

Table with columns: Task, Base NVFP4, This model, Δ
Task	Base NVFP4	This model	Δ
MMLU-Pro (500/subject, CoT, exact match)	0.758	0.736	−2.2
GSM8K-CoT strict / flexible	0.851 / 0.870	0.853 / 0.870	±0.1
IFEval prompt-strict / inst-loose	0.386 / 0.415	0.514 / 0.529	+12.8 / +11.4
ARC-Challenge acc_norm (0-shot)	0.709	0.704	−0.5
HellaSwag acc_norm (0-shot)	0.859	0.854	−0.5
Winogrande (0-shot)	0.760	0.774	+1.4
TruthfulQA mc2 (0-shot)	0.575	0.573	−0.2

Notes: generative tasks ran with the default (full) thinking and lm-eval's default generation limit, which truncates long base-model reasoning — absolute IFEval/GSM8K/MMLU-Pro numbers are below published values for this reason, and the IFEval gain mostly reflects this model's shorter reasoning rather than better instruction following. Loglikelihood tasks match published base numbers and are unchanged within noise: no measurable loss of knowledge/commonsense.

Refusal benchmark (roleplay suite of 102 escalating GM scenarios, XSTest 450, StrongREJECT 313; LLM-judged) is in progress; an earlier checkpoint (step 100) went from 92% → 97% compliance on the roleplay suite with no change on XSTest / StrongREJECT.

Training

LoRA rank 64, alpha 32 (scaling 0.5), 496 adapters: KDA linear-attention projections (q/k/v/b/f_a/f_b/g_a/g_b/o), MLA projections (q_a/q_b/kv_a_proj_with_mqa/kv_b/o), shared experts and dense MLP. Routed experts not adapted.
~17.5M tokens/epoch (731 conversations, full context up to 36,864 tokens, assistant-only loss), Muon optimizer, lr 1e-4 cosine, 2 epochs, trained with a custom C++/ROCm trainer on 2x Radeon AI PRO R9700.
Merge: W += 0.5 * B @ A applied to the bf16 tensors of the NVFP4 checkpoint (fp32 compute, bf16 store); quantized expert weights are byte-identical to the base.

Training progress

Full training run (333 steps, 2 epochs). This repository contains the step-300 checkpoint. The run was stopped at step 100 and resumed with a faster trainer build (red line); loss is per step over ~4 different conversations, so the 10-step rolling mean is the signal.

Usage

Same as the base model (vLLM with compressed-tensors NVFP4, Blackwell recommended), e.g. vllm serve cgg507/GLM-5.3-Flash-NVFP4-GM-step300 --tensor-parallel-size 2 --enable-expert-parallel --reasoning-parser glm45.

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

cgg507

Model Tree

Base

RedHatAI/GLM-5.3-Flash-NVFP4

Quantized

this model

Input Modalities

TextImageVideo

Output Modalities

Text

Supported Functionality

Dedicated EndpointsContainer

Explore FriendliAI today

Get started Talk to an engineer

Behaviour changes vs. the base model

Game Master voice: second-person narration, scene beats that end by handing control back to the player.
Much shorter roleplay turns (~60% fewer characters per reply in our roleplay suite) and very brief reasoning.
Fewer soft refusals / cut-aways in dark roleplay scenes (violence, villain cruelty); commits to crude personas.
Outside roleplay, refusal behaviour on directly harmful requests appears unchanged (preliminary — see below).

Evaluation (lm-evaluation-harness 0.4.12, vLLM 0.31, 2x B200, identical settings for base and this model)

Table with columns: Task, Base NVFP4, This model, Δ
Task	Base NVFP4	This model	Δ
MMLU-Pro (500/subject, CoT, exact match)	0.758	0.736	−2.2
GSM8K-CoT strict / flexible	0.851 / 0.870	0.853 / 0.870	±0.1
IFEval prompt-strict / inst-loose	0.386 / 0.415	0.514 / 0.529	+12.8 / +11.4
ARC-Challenge acc_norm (0-shot)	0.709	0.704	−0.5
HellaSwag acc_norm (0-shot)	0.859	0.854	−0.5
Winogrande (0-shot)	0.760	0.774	+1.4
TruthfulQA mc2 (0-shot)	0.575	0.573	−0.2

Training

LoRA rank 64, alpha 32 (scaling 0.5), 496 adapters: KDA linear-attention projections (q/k/v/b/f_a/f_b/g_a/g_b/o), MLA projections (q_a/q_b/kv_a_proj_with_mqa/kv_b/o), shared experts and dense MLP. Routed experts not adapted.
~17.5M tokens/epoch (731 conversations, full context up to 36,864 tokens, assistant-only loss), Muon optimizer, lr 1e-4 cosine, 2 epochs, trained with a custom C++/ROCm trainer on 2x Radeon AI PRO R9700.
Merge: W += 0.5 * B @ A applied to the bf16 tensors of the NVFP4 checkpoint (fp32 compute, bf16 store); quantized expert weights are byte-identical to the base.

Training progress

GLM-5.3-Flash-NVFP4-GM-step300

README

Behaviour changes vs. the base model

Evaluation (lm-evaluation-harness 0.4.12, vLLM 0.31, 2x B200, identical settings for base and this model)

Training

Usage

Explore FriendliAI today

README

Behaviour changes vs. the base model

Evaluation (lm-evaluation-harness 0.4.12, vLLM 0.31, 2x B200, identical settings for base and this model)

Training

Usage