GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Gyubeum
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Gyubeum
Model Tree
Qwen/Qwen3-VL-8B-Instruct
Input Modalities
Output Modalities
Supported Functionality
Gyubeum/UI_Genie_256k, bt/ split: 127,993 pairs (255,986 samples).Scalar reward signal for GUI agent training (PPO/GRPO, RLHF) or best-of-N selection. Unlike the discrete <|+|> / <|-|> classifier (SFT-128K), this model outputs a continuous, unbounded reward that can be used directly as a value signal.
vLLM's Qwen3VLForConditionalGeneration loader does not support the extra score.weight tensor, so use HuggingFace transformers directly.
python
import json, torchfrom safetensors import safe_openfrom huggingface_hub import hf_hub_downloadfrom transformers import AutoProcessor, Qwen3VLForConditionalGeneration MODEL_PATH = "Gyubeum/Qwen3-VL-8B-Instruct-UI-Genie-BT-256K"DEVICE = "cuda" model = Qwen3VLForConditionalGeneration.from_pretrained( MODEL_PATH, torch_dtype=torch.bfloat16, low_cpu_mem_usage=True,).to(DEVICE).eval()processor = AutoProcessor.from_pretrained(MODEL_PATH, max_pixels=1_048_576) # `score.weight` is not wired into the base CausalLM class, so load it separately.index = json.load(open(hf_hub_download(MODEL_PATH, "model.safetensors.index.json")))shard = hf_hub_download(MODEL_PATH, index["weight_map"]["score.weight"])with safe_open(shard, framework="pt") as f: w = f.get_tensor("score.weight")score_head = torch.nn.Linear(w.shape[1], w.shape[0], bias=False)with torch.no_grad(): score_head.weight.copy_(w)score_head = score_head.to(DEVICE, dtype=torch.bfloat16) # score = score_head(hidden_states[:, last_non_pad_index, :])| Field | Value |
|---|---|
| Base model | Qwen/Qwen3-VL-8B-Instruct + a freshly initialized score head |
| Training method | Bradley-Terry pairwise loss (LoRA, merged) |
| Training data | 127,993 preference pairs = 255,986 samples (Gyubeum/UI_Genie_256k bt/) |
| Epochs / steps | 1 epoch, 2,000 steps |
| Batch | 64 pairs per step (1 per GPU × 8 grad-accum × 8 GPUs) |
| Trainable | LoRA r=64, α=32, dropout 0.05 on all LLM linear layers + score head (174.6M params); vision tower and aligner frozen |
| Optimizer | AdamW, lr 1e-4, cosine schedule, 5% warmup, weight decay 0.1, grad clip 1.0 |
| Max length | 8,192 tokens (left truncation), max_pixels 1,048,576 |
| Precision / parallelism | bfloat16, DeepSpeed ZeRO-3, 8 GPUs |
| Architecture | Qwen3VLForConditionalGeneration + score.weight linear head |
| Task type | seq_cls (regression, num_labels=1) |
| Score | Last non-padding token hidden state → linear head → scalar |
Training code: train_rm/rm_main_scripts/bt/train_bt_256k.sh in AndroidFlux.
<|+|> / <|-|> classifier trained on the 127,976 single-label samples from 63,988 pairs.bibtex
@misc{qwen3technicalreport, title={Qwen3 Technical Report}, author={Qwen Team}, year={2025}, eprint={2505.09388}, archivePrefix={arXiv}, primaryClass={cs.CL},}import json, torchfrom safetensors import safe_openfrom huggingface_hub import hf_hub_downloadfrom transformers import AutoProcessor, Qwen3VLForConditionalGeneration MODEL_PATH = "Gyubeum/Qwen3-VL-8B-Instruct-UI-Genie-BT-256K"DEVICE = "cuda" model = Qwen3VLForConditionalGeneration.from_pretrained( MODEL_PATH, torch_dtype=torch.bfloat16, low_cpu_mem_usage=True,).to(DEVICE).eval()processor = AutoProcessor.from_pretrained(MODEL_PATH, max_pixels=1_048_576) # `score.weight` is not wired into the base CausalLM class, so load it separately.index = json.load(open(hf_hub_download(MODEL_PATH, "model.safetensors.index.json")))shard = hf_hub_download(MODEL_PATH, index["weight_map"]["score.weight"])with safe_open(shard, framework="pt") as f: w = f.get_tensor("score.weight")score_head = torch.nn.Linear(w.shape[1], w.shape[0], bias=False)with torch.no_grad(): score_head.weight.copy_(w)score_head = score_head.to(DEVICE, dtype=torch.bfloat16) # score = score_head(hidden_states[:, last_non_pad_index, :])@misc{qwen3technicalreport, title={Qwen3 Technical Report}, author={Qwen Team}, year={2025}, eprint={2505.09388}, archivePrefix={arXiv}, primaryClass={cs.CL},}