GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
fcpig
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
fcpig
Model Tree
Qwen/Qwen3.8-27B
Input Modalities
Output Modalities
Supported Functionality
| 项目 | 内容 |
|---|---|
| 基座模型 | Qwen/Qwen3.8-27B（原生多模态稠密模型，262K 上下文，Apache-2.0） |
| 微调方式 | LoRA（QLoRA 4-bit 训练），框架 LLaMA-Factory |
| 训练阶段 | Supervised Fine-Tuning（SFT） |
| 训练数据 | lure_v2 数据集，共 246 条多模态（图像 + 文本）指令样本 |
| 可训练参数 | 58,363,904（占全部 27.4B 参数的 0.21%） |
| 权重文件 | adapter_model.safetensors（222.8 MB） |
| 训练日期 | 2026-09-03 |
| 超参数 | 值 |
|---|---|
| learning_rate | 5e-05 |
| lr_scheduler_type | cosine |
| num_train_epochs | 6.0 |
| per_device_train_batch_size | 2 |
| gradient_accumulation_steps | 2（等效总 batch size = 4） |
| cutoff_len | 2048 |
| lora_rank | 8 |
| lora_alpha | 16 |
| lora_dropout | 0.05 |
python
import torchfrom transformers import AutoProcessor, AutoModelForImageTextToText, BitsAndBytesConfigfrom peft import PeftModel BASE = "Qwen/Qwen3.8-27B"ADAPTER = "fcpig/Lure" # 本仓库，或本地路径 ./train_0903 # 与训练时一致：4-bit 量化加载，降低显存占用（约 20GB 级显卡可运行）bnb = BitsAndBytesConfig( load_in_4bit=True, bnb_4bit_quant_type="nf4", bnb_4bit_compute_dtype=torch.bfloat16, bnb_4bit_use_double_quant=True,) model = AutoModelForImageTextToText.from_pretrained( BASE, quantization_config=bnb, device_map="auto", trust_remote_code=True,)model = PeftModel.from_pretrained(model, ADAPTER)processor = AutoProcessor.from_pretrained(ADAPTER) # 仓库内含 processor/tokenizer messages = [{ "role": "user", "content": [ {"type": "image", "image": "file://你的图片路径.jpg"}, {"type": "text", "text": "描述这张图片。"}, ],}]text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)inputs = processor(text=[text], images=["你的图片路径.jpg"], return_tensors="pt").to(model.device)out = model.generate(**inputs, max_new_tokens=512)print(processor.batch_decode(out, skip_special_tokens=True)[0])bash
llamafactory-cli chat \ --model_name_or_path Qwen/Qwen3.8-27B \ --adapter_name_or_path fcpig/Lure \ --template qwen3_8 \ --finetuning_type lorabash
llamafactory-cli export \ --model_name_or_path Qwen/Qwen3.8-27B \ --adapter_name_or_path fcpig/Lure \ --template qwen3_8 \ --export_dir exports/train_0903-merged \ --export_size 20qwen3_5/Qwen3.8 架构）：bash
pip install "transformers>=5.2.0" peft==0.18.1 bitsandbytesgit clone https://github.com/hiyouga/LLaMA-Factory.gitcd LLaMA-Factory && pip install -e ".[torch,metrics]"lure_v2（246 条多模态样本，sharegpt 格式，含 train.jsonl 与对应图像文件），在 data/dataset_info.json 中注册：json
{ "lure_v2": { "file_name": "train.jsonl", "formatting": "sharegpt", "columns": { "messages": "conversations" }, "tags": { "image": "image" } }}train_0903.yaml：yaml
model_name_or_path: Qwen/Qwen3.8-27Btrust_remote_code: truestage: sftfinetuning_type: loratemplate: qwen3_8enable_thinking: truedo_train: truedataset: lure_v2dataset_dir: datacutoff_len: 2048max_samples: 100000preprocessing_num_workers: 16per_device_train_batch_size: 2gradient_accumulation_steps: 2learning_rate: 5.0e-5num_train_epochs: 6.0lr_scheduler_type: cosinewarmup_steps: 0max_grad_norm: 1.0optim: adamw_torchbf16: truequantization_bit: 4quantization_method: bnbdouble_quantization: truefreeze_vision_tower: truefreeze_multi_modal_projector: truelora_rank: 8lora_alpha: 16lora_dropout: 0.05lora_target: allimage_max_pixels: 589824image_min_pixels: 1024video_max_pixels: 65536video_min_pixels: 256logging_steps: 5save_steps: 100plot_loss: truereport_to: noneseed: 42output_dir: saves/train_0903bash
llamafactory-cli train train_0903.yaml固定 seed: 42、相同软硬件环境下可基本复现训练曲线（量化训练存在少量非确定性，最终 loss 应收敛至 0.25 左右）。
| 文件 | 说明 |
|---|---|
adapter_model.safetensors | LoRA 微调权重（核心文件） |
adapter_config.json | LoRA 配置（基座模型已指向 Qwen/Qwen3.8-27B，可从 Hub 直接加载） |
tokenizer.json / tokenizer_config.json / chat_template.jinja | 分词器与对话模板 |
processor_config.json | 多模态处理器配置（图像/视频预处理参数） |
training_args.yaml |
注：训练过程中的中间检查点（checkpoint-100/200/300/372）与最终权重内容一致（checkpoint-372 即第 372 步 = 6 epoch 结束），未随仓库上传；training_args.bin、WebUI 日志等运行时文件亦未包含。
model_name_or_path: Qwen/Qwen3.8-27Btrust_remote_code: truestage: sftfinetuning_type: loratemplate: qwen3_8enable_thinking: truedo_train: truedataset: lure_v2dataset_dir: datacutoff_len: 2048max_samples: 100000preprocessing_num_workers: 16per_device_train_batch_size: 2gradient_accumulation_steps: 2learning_rate: 5.0e-5num_train_epochs: 6.0lr_scheduler_type: cosinewarmup_steps: 0max_grad_norm: 1.0optim: adamw_torchbf16: truequantization_bit: 4quantization_method: bnbdouble_quantization: truefreeze_vision_tower: truefreeze_multi_modal_projector: truelora_rank: 8lora_alpha: 16lora_dropout: 0.05lora_target: allimage_max_pixels: 589824image_min_pixels: 1024video_max_pixels: 65536video_min_pixels: 256logging_steps: 5save_steps: 100plot_loss: truereport_to: noneseed: 42output_dir: saves/train_0903| lora_target | all（含 q/k/v/o_proj、gate/up/down_proj 及 GDN 线性注意力 in_proj/out_proj 等） |
| quantization_bit / method | 4 / bitsandbytes（QLoRA，double_quantization: true） |
| freeze_vision_tower | true（视觉塔冻结） |
| freeze_multi_modal_projector | true |
| template | qwen3_8 |
| enable_thinking | true |
| warmup_steps | 0 |
| max_grad_norm | 1.0 |
| optimizer | adamw_torch（betas=0.9,0.999，eps=1e-08） |
| bf16 | true |
| seed | 42 |
| image_max_pixels / min_pixels | 768×768 / 32×32 |
| video_max_pixels / min_pixels | 256×256 / 16×16 |
| 原始训练参数存档（LLaMA-Factory 导出） |
train_results.json / trainer_state.json | 训练指标与完整日志记录 |
training_loss.png | 训练损失曲线 |