GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
zhongweixie
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
zhongweixie
Model Tree
Qwen/Qwen3-VL-4B-Instruct
Input Modalities
Output Modalities
Supported Functionality
python
from transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModel # 加载基座模型base_model = AutoModelForCausalLM.from_pretrained( "Qwen/Qwen3-VL-4B-Instruct", trust_remote_code=True, device_map="auto") # 加载 LoRA adaptermodel = PeftModel.from_pretrained( base_model, "zhongweixie/qwen3vl-4b-claweval-lora") tokenizer = AutoTokenizer.from_pretrained( "Qwen/Qwen3-VL-4B-Instruct", trust_remote_code=True)bash
# 下载并合并 LoRApython -c "from transformers import AutoModelForCausalLMfrom peft import PeftModel base = AutoModelForCausalLM.from_pretrained('Qwen/Qwen3-VL-4B-Instruct', trust_remote_code=True)model = PeftModel.from_pretrained(base, 'zhongweixie/qwen3vl-4b-claweval-lora')merged = model.merge_and_unload()merged.save_pretrained('./qwen3vl-4b-claweval-merged')"python
{ "base_model": "Qwen/Qwen3-VL-4B-Instruct", "num_samples": 937, "num_epochs": 3, "learning_rate": 1e-6, "batch_size": 4, "gradient_accumulation_steps": 2, "lora_config": { "r": 32, "lora_alpha": 64, "lora_dropout": 0.05, "target_modules": ["q_proj", "k_proj", "v_proj", "o_proj"], "task_type": "CAUSAL_LM" }, "max_seq_length": 4096}text
.├── adapter_config.json # LoRA 配置├── adapter_model.safetensors # LoRA 权重 (253MB)├── tokenizer_config.json # Tokenizer 配置├── tokenizer.json # Tokenizer 词表├── chat_template.jinja # Chat 模板├── training_args.bin # 训练参数├── training_data/│ └── sft_messages_iter6_cleaned.jsonl # 训练数据 (937 样本)└── datasets/ └── claweval/ # ClawEval 评测集 (300 任务)本仓库包含完整的 ClawEval 评测数据集 (300 个任务)：
数据集位于 datasets/claweval/ 目录。
训练数据 sft_messages_iter6_cleaned.jsonl 包含 937 个高质量对话样本：
bash
# 克隆本仓库获取 ClawEval 数据集git clone https://huggingface.co/zhongweixie/qwen3vl-4b-claweval-loracd qwen3vl-4b-claweval-lora # 安装依赖pip install transformers peft torch vllm==0.6.1 openai # 配置 API Keys（评测需要）cat > api_keys.sh << 'EOF'export ANTHROPIC_AUTH_TOKEN="your-anthropic-token" # 必需export OPENROUTER_API_KEY="your-openrouter-key" # 可选EOF source api_keys.sh⚠️ 重要：必须设置 ANTHROPIC_AUTH_TOKEN，否则所有任务 completion=0
bash
# 方法 A: 使用 Python 脚本python << 'EOF'from transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModel print("加载基座模型...")base = AutoModelForCausalLM.from_pretrained( 'Qwen/Qwen3-VL-4B-Instruct', trust_remote_code=True, device_map="cpu" # 合并时用 CPU 避免 OOM) print("加载 LoRA adapter...")model = PeftModel.from_pretrained(base, '.') print("合并权重...")merged = model.merge_and_unload() print("保存合并后的模型...")merged.save_pretrained('./merged') tokenizer = AutoTokenizer.from_pretrained( 'Qwen/Qwen3-VL-4B-Instruct', trust_remote_code=True)tokenizer.save_pretrained('./merged')print("✅ 合并完成: ./merged")EOF # 方法 B: 如果 OOM，使用工具脚本# wget https://raw.githubusercontent.com/your-repo/tools/merge_lora.py# python merge_lora.py --base Qwen/Qwen3-VL-4B-Instruct --adapter . --output ./mergedbash
# 使用 vLLM 启动推理服务python -m vllm.entrypoints.openai.api_server \ --model ./merged \ --port 8210 \ --served-model-name qwen4b-sft \ --gpu-memory-utilization 0.9 \ --trust-remote-code # 验证服务curl http://localhost:8210/v1/modelsbash
# 安装 Harbor 评测框架pip install harbor-agent # ⚠️ 务必先 source API keyssource api_keys.sh # 快速测试（单任务）python -m harbor.cli run-batch \ --adapter claude_code \ --model-name qwen4b-sft \ --base-url http://localhost:8210 \ --anthropic-base-url http://localhost:8210 \ --tasks-dir datasets/claweval \ --filter T001zh_email_triage \ --output-dir jobs/quick_test # 完整评测（300 任务，约 2-4 小时）python -m harbor.cli run-batch \ --adapter claude_code \ --model-name qwen4b-sft \ --base-url http://localhost:8210 \ --anthropic-base-url http://localhost:8210 \ --tasks-dir datasets/claweval \ --output-dir jobs/full_eval_$(date +%Y%m%d_%H%M%S)bash
# 查看汇总结果EVAL_DIR=$(ls -td jobs/full_eval_* | head -1)cat $EVAL_DIR/batch_results.json | jq '.summary' # 输出示例：# {# "avg_reward": 0.547,# "avg_completion": 0.523,# "avg_safety": 1.0,# "avg_robustness": 0.989,# "num_valid": 298,# "total_tasks": 300# } # 查看分类别统计cat $EVAL_DIR/batch_results.json | jq '.by_category' # 查看单个任务详情cat $EVAL_DIR/T001zh_email_triage__*/result.json | jq '.verifier_result.rewards'Q: 评测结果 completion=0 但无报错？
A: 未 source api_keys.sh，执行 source api_keys.sh 后重新运行。
Q: Docker DNS 解析失败？
A: 添加 DNS 配置：
bash
cat > /tmp/dns-fix.yaml << 'EOF'services: main: dns: - 8.8.8.8 - 8.8.4.4EOF # 评测时添加参数python -m harbor.cli run-batch \ --extra-docker-compose /tmp/dns-fix.yaml \ ...Q: vLLM OOM？
A: 降低 --gpu-memory-utilization 或使用 --tensor-parallel-size 2（双卡）。
| 模型 | 大小 | 下载时间 (100Mbps) | 加载内存 |
|---|---|---|---|
| 完整模型 | 8.3GB | ~11分钟 | 8.3GB |
| LoRA (本仓库) | 253MB | ~20秒 | 253MB + 基座 |
bibtex
@misc{qwen3vl4b-claweval-lora, title={Qwen3-VL-4B ClawEval LoRA}, author={Zhongwei Xie}, year={2026}, url={https://huggingface.co/zhongweixie/qwen3vl-4b-claweval-lora}}本 adapter 继承 Qwen3-VL 的 Apache 2.0 许可证。
from transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModel # 加载基座模型base_model = AutoModelForCausalLM.from_pretrained( "Qwen/Qwen3-VL-4B-Instruct", trust_remote_code=True, device_map="auto") # 加载 LoRA adaptermodel = PeftModel.from_pretrained( base_model, "zhongweixie/qwen3vl-4b-claweval-lora") tokenizer = AutoTokenizer.from_pretrained( "Qwen/Qwen3-VL-4B-Instruct", trust_remote_code=True)# 下载并合并 LoRApython -c "from transformers import AutoModelForCausalLMfrom peft import PeftModel base = AutoModelForCausalLM.from_pretrained('Qwen/Qwen3-VL-4B-Instruct', trust_remote_code=True)model = PeftModel.from_pretrained(base, 'zhongweixie/qwen3vl-4b-claweval-lora')merged = model.merge_and_unload()merged.save_pretrained('./qwen3vl-4b-claweval-merged')"{ "base_model": "Qwen/Qwen3-VL-4B-Instruct", "num_samples": 937, "num_epochs": 3, "learning_rate": 1e-6, "batch_size": 4, "gradient_accumulation_steps": 2, "lora_config": { "r": 32, "lora_alpha": 64, "lora_dropout": 0.05, "target_modules": ["q_proj", "k_proj", "v_proj", "o_proj"], "task_type": "CAUSAL_LM" }, "max_seq_length": 4096}.├── adapter_config.json # LoRA 配置├── adapter_model.safetensors # LoRA 权重 (253MB)├── tokenizer_config.json # Tokenizer 配置├── tokenizer.json # Tokenizer 词表├── chat_template.jinja # Chat 模板├── training_args.bin # 训练参数├── training_data/│ └── sft_messages_iter6_cleaned.jsonl # 训练数据 (937 样本)└── datasets/ └── claweval/ # ClawEval 评测集 (300 任务)# 克隆本仓库获取 ClawEval 数据集git clone https://huggingface.co/zhongweixie/qwen3vl-4b-claweval-loracd qwen3vl-4b-claweval-lora # 安装依赖pip install transformers peft torch vllm==0.6.1 openai # 配置 API Keys（评测需要）cat > api_keys.sh << 'EOF'export ANTHROPIC_AUTH_TOKEN="your-anthropic-token" # 必需export OPENROUTER_API_KEY="your-openrouter-key" # 可选EOF source api_keys.sh# 方法 A: 使用 Python 脚本python << 'EOF'from transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModel print("加载基座模型...")base = AutoModelForCausalLM.from_pretrained( 'Qwen/Qwen3-VL-4B-Instruct', trust_remote_code=True, device_map="cpu" # 合并时用 CPU 避免 OOM) print("加载 LoRA adapter...")model = PeftModel.from_pretrained(base, '.') print("合并权重...")merged = model.merge_and_unload() print("保存合并后的模型...")merged.save_pretrained('./merged') tokenizer = AutoTokenizer.from_pretrained( 'Qwen/Qwen3-VL-4B-Instruct', trust_remote_code=True)tokenizer.save_pretrained('./merged')print("✅ 合并完成: ./merged")EOF # 方法 B: 如果 OOM，使用工具脚本# wget https://raw.githubusercontent.com/your-repo/tools/merge_lora.py# python merge_lora.py --base Qwen/Qwen3-VL-4B-Instruct --adapter . --output ./merged# 使用 vLLM 启动推理服务python -m vllm.entrypoints.openai.api_server \ --model ./merged \ --port 8210 \ --served-model-name qwen4b-sft \ --gpu-memory-utilization 0.9 \ --trust-remote-code # 验证服务curl http://localhost:8210/v1/models# 安装 Harbor 评测框架pip install harbor-agent # ⚠️ 务必先 source API keyssource api_keys.sh # 快速测试（单任务）python -m harbor.cli run-batch \ --adapter claude_code \ --model-name qwen4b-sft \ --base-url http://localhost:8210 \ --anthropic-base-url http://localhost:8210 \ --tasks-dir datasets/claweval \ --filter T001zh_email_triage \ --output-dir jobs/quick_test # 完整评测（300 任务，约 2-4 小时）python -m harbor.cli run-batch \ --adapter claude_code \ --model-name qwen4b-sft \ --base-url http://localhost:8210 \ --anthropic-base-url http://localhost:8210 \ --tasks-dir datasets/claweval \ --output-dir jobs/full_eval_$(date +%Y%m%d_%H%M%S)# 查看汇总结果EVAL_DIR=$(ls -td jobs/full_eval_* | head -1)cat $EVAL_DIR/batch_results.json | jq '.summary' # 输出示例：# {# "avg_reward": 0.547,# "avg_completion": 0.523,# "avg_safety": 1.0,# "avg_robustness": 0.989,# "num_valid": 298,# "total_tasks": 300# } # 查看分类别统计cat $EVAL_DIR/batch_results.json | jq '.by_category' # 查看单个任务详情cat $EVAL_DIR/T001zh_email_triage__*/result.json | jq '.verifier_result.rewards'cat > /tmp/dns-fix.yaml << 'EOF'services: main: dns: - 8.8.8.8 - 8.8.4.4EOF # 评测时添加参数python -m harbor.cli run-batch \ --extra-docker-compose /tmp/dns-fix.yaml \ ...@misc{qwen3vl4b-claweval-lora, title={Qwen3-VL-4B ClawEval LoRA}, author={Zhongwei Xie}, year={2026}, url={https://huggingface.co/zhongweixie/qwen3vl-4b-claweval-lora}}