GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
SriGoli0123
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
SriGoli0123
Model Tree
Qwen/Qwen3-4B-Base
Input Modalities
Output Modalities
Supported Functionality
The root model-*.safetensors files contain the merged actor weights from training step 204 of 204. They load directly with Transformers:
python
import torchfrom transformers import AutoModelForCausalLM, AutoTokenizer model_id = "SriGoli0123/Qwen3-4B-Base-GRPO-641644"tokenizer = AutoTokenizer.from_pretrained(model_id)model = AutoModelForCausalLM.from_pretrained( model_id, torch_dtype=torch.bfloat16, device_map="auto",)The merged export contains 4,022,468,096 bfloat16 parameters across 398 tensors. The final deterministic DAPO Math validation accuracy was 0.341 on 1,000 examples at step 204 (best during training: 0.343 at step 130).
Qwen/Qwen3-4B-Base906bfd4b4dc7f14ee4320094d8b41684abff8539a100/ioher-mistake-judge, commit d169b69)actor.ioh_sft_coef=0)data.train_batch_size): 64 promptsrollout.n): 8actor.ppo_mini_batch_size): 16 prompts (4 optimizer steps per training step)actor.ppo_micro_batch_size_per_gpu): 4rollout.log_prob_micro_batch_size_per_gpu): 4actor.use_dynamic_bsz, rollout.log_prob_use_dynamic_bsz): enabledactor.ppo_max_token_len_per_gpu, rollout.log_prob_max_token_len_per_gpu): 24,5761e-6[0.9, 0.999]torch.compile: disabled-1)This is a research checkpoint. Validation accuracy plateaued around 0.31-0.34 after roughly step 70; differences of about ±1.5 points are within noise on the 1,000-example validation set. Evaluate the model on the intended benchmark and decoding configuration before downstream use.
actor.ppo_epochs): 1data.max_prompt_length): 2,048 tokensdata.max_response_length): 4,096 tokens