GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
ertghiu256
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
ertghiu256
Model Tree
ertghiu256/Qwen3.5-2b-ReMix
Input Modalities
Output Modalities
Supported Functionality
ertghiu256/Qwen3.5-2b-ReMix (a fine‑tune of Qwen3.5‑2B)loss_type = "dr_grpo"GRPOTrainerTrained with these GRPOConfig goodies:
yaml
learning_rate: 4e-6adam_beta1: 0.9adam_beta2: 0.99weight_decay: 0.1warmup_ratio: 0.1lr_scheduler_type: cosineoptim: adamw_8bitlogging_steps: 1per_device_train_batch_size: 1gradient_accumulation_steps: 1num_generations: 2max_prompt_length: 1024max_completion_length: 1024max_steps: 100max_grad_norm: 0.1importance_sampling_level: sequencemask_truncated_completions: Falseloss_type: dr_grpoTraining ran on Unsloth’s Vision GRPO notebook. With just 100 steps of RL.
The model expects this structure:
python
REASONING_START = "<think>"REASONING_END = "</think>"SOLUTION_START = "<answer>"SOLUTION_END = "</answer>"Prompt in training (image + text):
python
{ "role": "user", "content": [ {"type": "image"}, {"type": "text", "text": "Your question here. Also first provide your reasoning... and then your final answer between <answer> and (put a single float here) </answer>"} ]}The model learned to spit out reasoning inside <think> and the final numeric answer inside <answer>. Short & clean. ✨
| Parameter | Value |
|---|---|
temperature | 0.4 |
top_k | 30 |
repeat_penalty | 1.1 |
| Parameter | Value |
|---|---|
temperature | 0.0 – 0.1 |
top_k | 60 |
repeat_penalty | 1.2 |
python
from unsloth import FastVisionModelfrom transformers import AutoProcessorimport torch model, tokenizer = FastVisionModel.from_pretrained( "ertghiu256/Qwen3.5-2b-ReMix-Vision-GRPO", max_seq_length=1024, # 👈 stop from overthinking! load_in_4bit=False, fast_inference=False,) processor = AutoProcessor.from_pretrained("ertghiu256/Qwen3.5-2b-ReMix-Vision-GRPO") messages = [ { "role": "user", "content": [ {"type": "image"}, {"type": "text", "text": "What's the area? Also first provide your reasoning... and then your final answer between <answer> and (put a single float here) </answer>"} ] }] inputs = processor.apply_chat_template( messages, add_generation_prompt=True, tokenize=True, return_tensors="pt", return_dict=True,)inputs = {k: v.to(model.device) for k, v in inputs.items()} outputs = model.generate( **inputs, max_new_tokens=1024, temperature=0.4, top_k=30, repetition_penalty=1.1, do_sample=True,)response = processor.decode(outputs[0], skip_special_tokens=False)print(response)This qwen3_5 model was trained 2x faster with Unsloth and Huggingface's TRL library.
REASONING_START = "<think>"REASONING_END = "</think>"SOLUTION_START = "<answer>"SOLUTION_END = "</answer>"{ "role": "user", "content": [ {"type": "image"}, {"type": "text", "text": "Your question here. Also first provide your reasoning... and then your final answer between <answer> and (put a single float here) </answer>"} ]}from unsloth import FastVisionModelfrom transformers import AutoProcessorimport torch model, tokenizer = FastVisionModel.from_pretrained( "ertghiu256/Qwen3.5-2b-ReMix-Vision-GRPO", max_seq_length=1024, # 👈 stop from overthinking! load_in_4bit=False, fast_inference=False,) processor = AutoProcessor.from_pretrained("ertghiu256/Qwen3.5-2b-ReMix-Vision-GRPO") messages = [ { "role": "user", "content": [ {"type": "image"}, {"type": "text", "text": "What's the area? Also first provide your reasoning... and then your final answer between <answer> and (put a single float here) </answer>"} ] }] inputs = processor.apply_chat_template( messages, add_generation_prompt=True, tokenize=True, return_tensors="pt", return_dict=True,)inputs = {k: v.to(model.device) for k, v in inputs.items()} outputs = model.generate( **inputs, max_new_tokens=1024, temperature=0.4, top_k=30, repetition_penalty=1.1, do_sample=True,)response = processor.decode(outputs[0], skip_special_tokens=False)print(response)learning_rate: 4e-6adam_beta1: 0.9adam_beta2: 0.99weight_decay: 0.1warmup_ratio: 0.1lr_scheduler_type: cosineoptim: adamw_8bitlogging_steps: 1per_device_train_batch_size: 1gradient_accumulation_steps: 1num_generations: 2max_prompt_length: 1024max_completion_length: 1024max_steps: 100max_grad_norm: 0.1importance_sampling_level: sequencemask_truncated_completions: Falseloss_type: dr_grpo