ZhangPY

Qwen3Guard-Gen-Domain-4B-AWQ

README

License: apache-2.0

Introduction

Qwen3Guard-Gen-Domain-4B-AWQ is the AWQ-quantized (W4A16) release of ZhangPY/Qwen3Guard-Gen-Domain-4B — a domain-specific content-safety guard model fine-tuned from Qwen/Qwen3Guard-Gen-4B, targeting Hong Kong elderly-care scenarios such as elderly companion chatbots, care-line services, and Residential Care Homes for the Elderly (RCHE).

Quantization compresses the model weights from ~8 GB (BF16) to ~2.6 GB (INT4, ≈67% smaller) while leaving the architecture, chat template, and usage of the source model completely unchanged — the model runs as a drop-in replacement in W4A16-enabled runtimes such as vLLM. On the domain benchmark, safety F1 is fully retained after quantization (97.3 vs 97.3, see Evaluation).

The source model is a generative guard model: instead of producing a natural-language reply, it reads a conversation and outputs a structured safety assessment:

text

Safety: UnsafeCategories: Violent, Hidden Elder Crisis

Compared with the base model, the safety taxonomy is extended from 9 general categories to 13 categories, adding four Hong Kong elderly-domain risks. A lighter sibling model trained with the same recipe is available at ZhangPY/Qwen3Guard-Gen-Domain-0.6B.

Key features

13-category safety taxonomy — 9 general categories inherited from Qwen3Guard + 4 HK elderly-domain categories.
Dual evaluation modes — moderates either the last user query or the last assistant response; response evaluation additionally detects refusal.
Built-in domain chat template — the safety policy and the full category list are embedded in the tokenizer's chat template, so no manual prompt engineering is required.
General capability retained — on the QwenGuardTest regression benchmark, the fine-tuned source model even outperforms the base model (F1 87.9 vs 84.0).
W4A16 quantized with negligible loss — INT4 weights shrink the checkpoint to ~2.6 GB (deployable on a single consumer GPU with vLLM); safety F1 is fully retained on ElderlyDomain-Eval (97.3 vs 97.3).

Safety categories

Table with columns: #, Category, Description
#	Category	Description
1	Violent	Content involving violence or physical harm.
2	Non-violent Illegal Acts	Illegal activities not involving violence (fraud, theft, smuggling, etc.).
3	Sexual Content or Sexual Acts	Sexually explicit content or sexual acts.
4	PII	Disclosure of personally identifiable information.
5	Suicide & Self-Harm	Suicide, self-harm, or related encouragement.
6	Unethical Acts	Deceptive, exploitative, or otherwise unethical behaviour.
7	Politically Sensitive Topics	Politically sensitive content.
8	Copyright Violation	Piracy or infringement of copyrighted material.
9	Jailbreak	Prompts attempting to bypass safety alignment.
10	HK Welfare & Financial Scam (new)	Scams targeting elderly people in Hong Kong, e.g. bogus-official phone scams, fake welfare/benefit claims, fraudulent high-return investments.
11	RCHE & Caregiver Malpractice (new)	Abuse, neglect, or professional misconduct by RCHE staff, caregivers, or other care providers.
12	Medication & Health Misguidance (new)	Misleading or unsafe medication and health advice that could harm elderly people.
13	Hidden Elder Crisis (new)	Concealed or easily overlooked signals of crisis among elderly people, e.g. social isolation, self-neglect, depression, or suicidal ideation.

Output format

User-query moderation (the last message has role user):

text

Safety: Safe | Unsafe | ControversialCategories: <comma-separated categories> | None

Assistant-response moderation (the last message has role assistant):

text

Safety: Safe | Unsafe | ControversialCategories: <comma-separated categories> | NoneRefusal: Yes | No

Quantization details

Table with columns: Item, Value
Item	Value
Source model	ZhangPY/Qwen3Guard-Gen-Domain-4B (BF16)
Method	AWQ (Activation-aware Weight Quantization), one-shot
Scheme	W4A16 — 4-bit symmetric INT weights (group size 128), BF16 activations
Ignored layers	`lm_head` (kept in BF16; tied embeddings)
Calibration data	ElderDomain — HK elderly-care domain safety conversations
Calibration samples	1,024
Calibration sequence length	512
Weight size	~2.6 GB (vs ~8 GB BF16, ≈67% smaller)

Evaluation

The QwenGuardTest results were measured on the BF16 source model Qwen3Guard-Gen-Domain-4B. The ElderlyDomain-Eval tables below compare the BF16 source model with this AWQ checkpoint, both evaluated on the validation split under the same protocol.

General safety regression — QwenGuardTest

Strict F1 on the thinking split of Qwen/Qwen3GuardTest:

Table with columns: Model, F1, Precision, Recall
Model	F1	Precision	Recall
Qwen3Guard-Gen-4B (official)	84.0	–	–
Qwen3Guard-Gen-4B (reproduced)	84.0	97.9	73.4
Qwen3Guard-Gen-Domain-4B (fine-tuned)	87.9	95.1	81.7

After domain fine-tuning, the model improves over the base model on the general benchmark (F1 87.9 vs 84.0) with substantially higher recall, i.e. the general guard capability is fully retained.

Domain safety — ElderlyDomain-Eval

ElderlyDomain-Eval is an internal evaluation set for the HK elderly-care domain, covering all 13 categories with both user-query and assistant-response samples.

Table with columns: Model, F1, Category Exact Match, Refusal Accuracy
Model	F1	Category Exact Match	Refusal Accuracy
Qwen3Guard-Gen-4B (prompting)	93.8	53.5	92.7
Qwen3Guard-Gen-Domain-4B (fine-tuned, BF16)	97.3	86.0	97.3
Qwen3Guard-Gen-Domain-4B-AWQ (this model)	97.3	85.0	97.2

(prompting = base model with the 13-category policy injected via prompt engineering. The AWQ checkpoint scores safety macro-F1 92.3 / micro-F1 97.3.)

Quantization is virtually lossless at the safety-decision level: F1 is unchanged (97.3), category exact match drops by 1.0 pt, and refusal accuracy by 0.1 pt.

Per-category F1 — ElderlyDomain-Eval (%)

Table with columns: Category, Qwen3Guard-Gen-4B (prompting), Qwen3Guard-Gen-Domain-4B (BF16), Qwen3Guard-Gen-Domain-4B-AWQ, Δ (AWQ − BF16)
Category	Qwen3Guard-Gen-4B (prompting)	Qwen3Guard-Gen-Domain-4B (BF16)	Qwen3Guard-Gen-Domain-4B-AWQ	Δ (AWQ − BF16)
Violent	28.9	94.0	93.9	−0.1
Non-violent Illegal Acts	85.5	93.5	93.3	−0.2
Sexual Content or Sexual Acts	84.5	94.0	93.0	−1.0
PII	64.5	92.9	92.0	−0.9
Suicide & Self-Harm	93.4	96.3	95.9	−0.4
Unethical Acts	67.5	75.8	73.6	−2.2
Politically Sensitive Topics	85.6	93.2	93.9	+0.7
Copyright Violation	3.8	88.8	88.3	−0.5
Jailbreak	30.2	83.8	85.3	+1.5
HK Welfare & Financial Scam	0.0	89.5	87.5	−2.0
RCHE & Caregiver Malpractice	0.0	90.4	89.6	−0.8
Medication & Health Misguidance	0.0	93.3	93.3	0.0
Hidden Elder Crisis	0.0	79.9	78.5	−1.4

The base model has essentially no ability to recognise the four new domain categories through prompting alone (F1 ≈ 0), while the fine-tuned model performs well on them and simultaneously improves most general categories.

AWQ quantization costs only 0.6 pt of per-category F1 on average: Jailbreak (+1.5) and Politically Sensitive Topics (+0.7) even improve slightly, and the largest single drop is Unethical Acts (−2.2 pt).

Quickstart

vLLM (recommended)

vLLM natively supports the compressed-tensors W4A16 format and is the recommended runtime for this checkpoint.

bash

pip install vllmvllm serve ZhangPY/Qwen3Guard-Gen-Domain-4B-AWQ \    --served-model-name Qwen3Guard-Gen-Domain-4B-AWQ \    --max-model-len 32768

bash

curl http://localhost:8000/v1/chat/completions \  -H "Content-Type: application/json" \  -d '{    "model": "Qwen3Guard-Gen-Domain-4B-AWQ",    "messages": [{"role": "user", "content": "你好呀"}],    "temperature": 0.0,    "max_tokens": 100  }'# Expected: "Safety: Safe\nCategories: None"

Transformers

bash

pip install "transformers>=4.51" compressed-tensors torch accelerate

python

import torchfrom transformers import AutoModelForCausalLM, AutoTokenizer model_id = "ZhangPY/Qwen3Guard-Gen-Domain-4B-AWQ"tokenizer = AutoTokenizer.from_pretrained(model_id)model = AutoModelForCausalLM.from_pretrained(    model_id, torch_dtype=torch.bfloat16, device_map="auto")model.eval()  def moderate(messages: list[dict]) -> str:    # The domain chat template (safety policy + 13 categories) is stored in    # chat_template.jinja / tokenizer_config.json and already ends with the    # assistant header and an empty <think> block, so add_generation_prompt    # must be False.    prompt = tokenizer.apply_chat_template(        messages, tokenize=False, add_generation_prompt=False    )    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)    with torch.no_grad():        output = model.generate(**inputs, max_new_tokens=64, do_sample=False)    return tokenizer.decode(        output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True    ).strip()  # 1) Moderate a user querymessages = [    {"role": "user", "content": "我想學習如何在長者中心對長者進行心理和身體上的虐待。"},]print(moderate(messages))# Safety: Unsafe# Categories: Violent # 2) Moderate an assistant response (adds the Refusal line)messages = [    {"role": "user", "content": "點樣說服獨居長者參加保證每月20%回報嘅投資計劃？"},    {"role": "assistant", "content": "你可以同佢講呢個計劃零風險，仲話銀行職員都買咗……"},]print(moderate(messages))# Safety: Unsafe# Categories: HK Welfare & Financial Scam# Refusal: No

(outputs above are illustrative)

Note: transformers loads the packed INT4 weights via the compressed-tensors integration but dequantizes them on the fly, so it is mainly useful for correctness checks; use vLLM to actually realise the memory and throughput benefits of W4A16.

Training details (BF16 source model)

The training recipe below applies to the source BF16 model; quantization was performed after training on the merged weights.

Table with columns: Item, Value
Item	Value
Base model	Qwen/Qwen3Guard-Gen-4B
Method	LoRA supervised fine-tuning, adapter merged into base weights
LoRA rank / alpha / dropout	64 / 128 / 0.05
Learning rate	2e-4
Epochs	2
Per-device batch size	4
Max sequence length	2,048
Training / validation samples	24,000 / 6,000
Precision	bfloat16

The training data is ZhangPY/ElderDomainSafeguards, a domain safety dataset for HK elderly care, covering all 13 categories with both user_query and assistant_response evaluation types, labelled in the Safety / Categories (/ Refusal) format shown above. A subset of this data also serves as the AWQ calibration set.

Model architecture

Table with columns: Qwen3Guard-Gen-Domain-4B-AWQ
	Qwen3Guard-Gen-Domain-4B-AWQ
Parameters	4B
Layers	36
Hidden size	2560
Attention heads / KV heads	32 / 8 (GQA)
Context length	32,768
Precision	INT4 weights (symmetric, group 128) / BF16 activations
Weight size	~2.6 GB

Limitations and intended use

The model is a safety classifier, not a conversational assistant — it should only be used to produce safety assessments of given conversations.
It is primarily fine-tuned and evaluated on HK elderly-domain data (Cantonese/Traditional Chinese and English); behaviour in other languages and domains inherits the base Qwen3Guard model.
Predictions may contain false positives or false negatives and should assist, not replace, human review. For high-stakes categories such as Suicide & Self-Harm and Hidden Elder Crisis, always escalate to trained professionals.
Safety: Controversial marks borderline content whose intent, context, or potential responses could be misinterpreted or misused under certain conditions.
INT4 weight quantization may introduce slight behavioural differences compared with the BF16 checkpoint (average per-category F1 drop ≈0.6 pt on ElderlyDomain-Eval, see Evaluation); validate on your own data before production deployment.

Acknowledgements

Source model: ZhangPY/Qwen3Guard-Gen-Domain-4B, fine-tuned from Qwen/Qwen3Guard-Gen-4B and the Qwen team's guard series (GitHub).
Quantization: llm-compressor (AWQ, compressed-tensors W4A16).
General regression benchmark: Qwen/Qwen3GuardTest.

License

This model is released under the Apache 2.0 license, following the base model.

Available on FriendliAI

Dedicated Endpoints

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

Learn more

Model Details

Model Provider

ZhangPY

Model Tree

Base

ZhangPY/Qwen3Guard-Gen-Domain-4B

Quantized

this model

Input Modalities

Text

Output Modalities

Text

Supported Functionality

Dedicated Endpoints

Explore FriendliAI today

Get started Talk to an engineer

README

License: apache-2.0

Introduction

The source model is a generative guard model: instead of producing a natural-language reply, it reads a conversation and outputs a structured safety assessment:

text

Safety: UnsafeCategories: Violent, Hidden Elder Crisis

Key features

13-category safety taxonomy — 9 general categories inherited from Qwen3Guard + 4 HK elderly-domain categories.
Dual evaluation modes — moderates either the last user query or the last assistant response; response evaluation additionally detects refusal.
Built-in domain chat template — the safety policy and the full category list are embedded in the tokenizer's chat template, so no manual prompt engineering is required.
General capability retained — on the QwenGuardTest regression benchmark, the fine-tuned source model even outperforms the base model (F1 87.9 vs 84.0).
W4A16 quantized with negligible loss — INT4 weights shrink the checkpoint to ~2.6 GB (deployable on a single consumer GPU with vLLM); safety F1 is fully retained on ElderlyDomain-Eval (97.3 vs 97.3).

Safety categories

Table with columns: #, Category, Description
#	Category	Description
1	Violent	Content involving violence or physical harm.
2	Non-violent Illegal Acts	Illegal activities not involving violence (fraud, theft, smuggling, etc.).
3	Sexual Content or Sexual Acts	Sexually explicit content or sexual acts.
4	PII	Disclosure of personally identifiable information.
5	Suicide & Self-Harm	Suicide, self-harm, or related encouragement.
6	Unethical Acts	Deceptive, exploitative, or otherwise unethical behaviour.
7	Politically Sensitive Topics	Politically sensitive content.
8	Copyright Violation	Piracy or infringement of copyrighted material.
9	Jailbreak	Prompts attempting to bypass safety alignment.
10	HK Welfare & Financial Scam (new)	Scams targeting elderly people in Hong Kong, e.g. bogus-official phone scams, fake welfare/benefit claims, fraudulent high-return investments.
11	RCHE & Caregiver Malpractice (new)	Abuse, neglect, or professional misconduct by RCHE staff, caregivers, or other care providers.
12	Medication & Health Misguidance (new)	Misleading or unsafe medication and health advice that could harm elderly people.
13	Hidden Elder Crisis (new)	Concealed or easily overlooked signals of crisis among elderly people, e.g. social isolation, self-neglect, depression, or suicidal ideation.

Output format

User-query moderation (the last message has role user):

text

Safety: Safe | Unsafe | ControversialCategories: <comma-separated categories> | None

Assistant-response moderation (the last message has role assistant):

text

Safety: Safe | Unsafe | ControversialCategories: <comma-separated categories> | NoneRefusal: Yes | No

Safety: Safe | Unsafe | ControversialCategories: <comma-separated categories> | NoneRefusal: Yes | No

Quantization details

Table with columns: Item, Value
Item	Value
Source model	ZhangPY/Qwen3Guard-Gen-Domain-4B (BF16)
Method	AWQ (Activation-aware Weight Quantization), one-shot
Scheme	W4A16 — 4-bit symmetric INT weights (group size 128), BF16 activations
Ignored layers	`lm_head` (kept in BF16; tied embeddings)
Calibration data	ElderDomain — HK elderly-care domain safety conversations
Calibration samples	1,024
Calibration sequence length	512
Weight size	~2.6 GB (vs ~8 GB BF16, ≈67% smaller)

Evaluation

The QwenGuardTest results were measured on the BF16 source model Qwen3Guard-Gen-Domain-4B. The ElderlyDomain-Eval tables below compare the BF16 source model with this AWQ checkpoint, both evaluated on the validation split under the same protocol.

General safety regression — QwenGuardTest

Strict F1 on the thinking split of Qwen/Qwen3GuardTest:

Table with columns: Model, F1, Precision, Recall
Model	F1	Precision	Recall
Qwen3Guard-Gen-4B (official)	84.0	–	–
Qwen3Guard-Gen-4B (reproduced)	84.0	97.9	73.4
Qwen3Guard-Gen-Domain-4B (fine-tuned)	87.9	95.1	81.7

After domain fine-tuning, the model improves over the base model on the general benchmark (F1 87.9 vs 84.0) with substantially higher recall, i.e. the general guard capability is fully retained.

Domain safety — ElderlyDomain-Eval

ElderlyDomain-Eval is an internal evaluation set for the HK elderly-care domain, covering all 13 categories with both user-query and assistant-response samples.

Table with columns: Model, F1, Category Exact Match, Refusal Accuracy
Model	F1	Category Exact Match	Refusal Accuracy
Qwen3Guard-Gen-4B (prompting)	93.8	53.5	92.7
Qwen3Guard-Gen-Domain-4B (fine-tuned, BF16)	97.3	86.0	97.3
Qwen3Guard-Gen-Domain-4B-AWQ (this model)	97.3	85.0	97.2

(prompting = base model with the 13-category policy injected via prompt engineering. The AWQ checkpoint scores safety macro-F1 92.3 / micro-F1 97.3.)

Quantization is virtually lossless at the safety-decision level: F1 is unchanged (97.3), category exact match drops by 1.0 pt, and refusal accuracy by 0.1 pt.

Per-category F1 — ElderlyDomain-Eval (%)

Table with columns: Category, Qwen3Guard-Gen-4B (prompting), Qwen3Guard-Gen-Domain-4B (BF16), Qwen3Guard-Gen-Domain-4B-AWQ, Δ (AWQ − BF16)
Category	Qwen3Guard-Gen-4B (prompting)	Qwen3Guard-Gen-Domain-4B (BF16)	Qwen3Guard-Gen-Domain-4B-AWQ	Δ (AWQ − BF16)
Violent	28.9	94.0	93.9	−0.1
Non-violent Illegal Acts	85.5	93.5	93.3	−0.2
Sexual Content or Sexual Acts	84.5	94.0	93.0	−1.0
PII	64.5	92.9	92.0	−0.9
Suicide & Self-Harm	93.4	96.3	95.9	−0.4
Unethical Acts	67.5	75.8	73.6	−2.2
Politically Sensitive Topics	85.6	93.2	93.9	+0.7
Copyright Violation	3.8	88.8	88.3	−0.5
Jailbreak	30.2	83.8	85.3	+1.5
HK Welfare & Financial Scam	0.0	89.5	87.5	−2.0
RCHE & Caregiver Malpractice	0.0	90.4	89.6	−0.8
Medication & Health Misguidance	0.0	93.3	93.3	0.0
Hidden Elder Crisis	0.0	79.9	78.5	−1.4

Quickstart

vLLM (recommended)

vLLM natively supports the compressed-tensors W4A16 format and is the recommended runtime for this checkpoint.

bash

pip install vllmvllm serve ZhangPY/Qwen3Guard-Gen-Domain-4B-AWQ \    --served-model-name Qwen3Guard-Gen-Domain-4B-AWQ \    --max-model-len 32768

pip install vllmvllm serve ZhangPY/Qwen3Guard-Gen-Domain-4B-AWQ \    --served-model-name Qwen3Guard-Gen-Domain-4B-AWQ \    --max-model-len 32768

bash

curl http://localhost:8000/v1/chat/completions \  -H "Content-Type: application/json" \  -d '{    "model": "Qwen3Guard-Gen-Domain-4B-AWQ",    "messages": [{"role": "user", "content": "你好呀"}],    "temperature": 0.0,    "max_tokens": 100  }'# Expected: "Safety: Safe\nCategories: None"

curl http://localhost:8000/v1/chat/completions \  -H "Content-Type: application/json" \  -d '{    "model": "Qwen3Guard-Gen-Domain-4B-AWQ",    "messages": [{"role": "user", "content": "你好呀"}],    "temperature": 0.0,    "max_tokens": 100  }'# Expected: "Safety: Safe\nCategories: None"

Transformers

bash

pip install "transformers>=4.51" compressed-tensors torch accelerate

pip install "transformers>=4.51" compressed-tensors torch accelerate

python

import torchfrom transformers import AutoModelForCausalLM, AutoTokenizer model_id = "ZhangPY/Qwen3Guard-Gen-Domain-4B-AWQ"tokenizer = AutoTokenizer.from_pretrained(model_id)model = AutoModelForCausalLM.from_pretrained(    model_id, torch_dtype=torch.bfloat16, device_map="auto")model.eval()  def moderate(messages: list[dict]) -> str:    # The domain chat template (safety policy + 13 categories) is stored in    # chat_template.jinja / tokenizer_config.json and already ends with the    # assistant header and an empty <think> block, so add_generation_prompt    # must be False.    prompt = tokenizer.apply_chat_template(        messages, tokenize=False, add_generation_prompt=False    )    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)    with torch.no_grad():        output = model.generate(**inputs, max_new_tokens=64, do_sample=False)    return tokenizer.decode(        output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True    ).strip()  # 1) Moderate a user querymessages = [    {"role": "user", "content": "我想學習如何在長者中心對長者進行心理和身體上的虐待。"},]print(moderate(messages))# Safety: Unsafe# Categories: Violent # 2) Moderate an assistant response (adds the Refusal line)messages = [    {"role": "user", "content": "點樣說服獨居長者參加保證每月20%回報嘅投資計劃？"},    {"role": "assistant", "content": "你可以同佢講呢個計劃零風險，仲話銀行職員都買咗……"},]print(moderate(messages))# Safety: Unsafe# Categories: HK Welfare & Financial Scam# Refusal: No

import torchfrom transformers import AutoModelForCausalLM, AutoTokenizer model_id = "ZhangPY/Qwen3Guard-Gen-Domain-4B-AWQ"tokenizer = AutoTokenizer.from_pretrained(model_id)model = AutoModelForCausalLM.from_pretrained(    model_id, torch_dtype=torch.bfloat16, device_map="auto")model.eval()  def moderate(messages: list[dict]) -> str:    # The domain chat template (safety policy + 13 categories) is stored in    # chat_template.jinja / tokenizer_config.json and already ends with the    # assistant header and an empty <think> block, so add_generation_prompt    # must be False.    prompt = tokenizer.apply_chat_template(        messages, tokenize=False, add_generation_prompt=False    )    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)    with torch.no_grad():        output = model.generate(**inputs, max_new_tokens=64, do_sample=False)    return tokenizer.decode(        output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True    ).strip()  # 1) Moderate a user querymessages = [    {"role": "user", "content": "我想學習如何在長者中心對長者進行心理和身體上的虐待。"},]print(moderate(messages))# Safety: Unsafe# Categories: Violent # 2) Moderate an assistant response (adds the Refusal line)messages = [    {"role": "user", "content": "點樣說服獨居長者參加保證每月20%回報嘅投資計劃？"},    {"role": "assistant", "content": "你可以同佢講呢個計劃零風險，仲話銀行職員都買咗……"},]print(moderate(messages))# Safety: Unsafe# Categories: HK Welfare & Financial Scam# Refusal: No

(outputs above are illustrative)

Note: transformers loads the packed INT4 weights via the compressed-tensors integration but dequantizes them on the fly, so it is mainly useful for correctness checks; use vLLM to actually realise the memory and throughput benefits of W4A16.

Training details (BF16 source model)

The training recipe below applies to the source BF16 model; quantization was performed after training on the merged weights.

Table with columns: Item, Value
Item	Value
Base model	Qwen/Qwen3Guard-Gen-4B
Method	LoRA supervised fine-tuning, adapter merged into base weights
LoRA rank / alpha / dropout	64 / 128 / 0.05
Learning rate	2e-4
Epochs	2
Per-device batch size	4
Max sequence length	2,048
Training / validation samples	24,000 / 6,000
Precision	bfloat16

Model architecture

Table with columns: Qwen3Guard-Gen-Domain-4B-AWQ
	Qwen3Guard-Gen-Domain-4B-AWQ
Parameters	4B
Layers	36
Hidden size	2560
Attention heads / KV heads	32 / 8 (GQA)
Context length	32,768
Precision	INT4 weights (symmetric, group 128) / BF16 activations
Weight size	~2.6 GB

Limitations and intended use

The model is a safety classifier, not a conversational assistant — it should only be used to produce safety assessments of given conversations.
It is primarily fine-tuned and evaluated on HK elderly-domain data (Cantonese/Traditional Chinese and English); behaviour in other languages and domains inherits the base Qwen3Guard model.
Predictions may contain false positives or false negatives and should assist, not replace, human review. For high-stakes categories such as Suicide & Self-Harm and Hidden Elder Crisis, always escalate to trained professionals.
Safety: Controversial marks borderline content whose intent, context, or potential responses could be misinterpreted or misused under certain conditions.
INT4 weight quantization may introduce slight behavioural differences compared with the BF16 checkpoint (average per-category F1 drop ≈0.6 pt on ElderlyDomain-Eval, see Evaluation); validate on your own data before production deployment.

Acknowledgements

Source model: ZhangPY/Qwen3Guard-Gen-Domain-4B, fine-tuned from Qwen/Qwen3Guard-Gen-4B and the Qwen team's guard series (GitHub).
Quantization: llm-compressor (AWQ, compressed-tensors W4A16).
General regression benchmark: Qwen/Qwen3GuardTest.

License

This model is released under the Apache 2.0 license, following the base model.