GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
HongbangYuan
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
HongbangYuan
Model Tree
Qwen/Qwen3.5-4B
Input Modalities
Output Modalities
Supported Functionality
ae9ef19dfcf2f233a07081dd6d927993ce894441gold200_astra_v1; 200 environment trajectories, 8,423 step-level training examples, 19 SciWorld L2 training task categories.| Setting | Value |
|---|---|
| Framework | LLaMA-Factory |
| Base model | Qwen/Qwen3.5-4B |
| Method | Full SFT on text; vision tower and multimodal projector frozen |
| Epochs / optimizer steps | 2 / 66 |
| GPUs | 4 |
| Per-device batch / gradient accumulation | 1 / 64 |
| Effective global batch | 256 |
| Learning rate | 1e-5 |
| Scheduler / warm-up ratio | Cosine / 0.1 |
| Maximum sequence length | 8,192 |
| Weight decay | 0.01 |
| Seed | 0 |
| Precision | BF16 |
| Attention / distributed optimizer | SDPA / DeepSpeed ZeRO-3 |
| Chat template | qwen3_5; enable_thinking=true |
| Packing / train_on_prompt | false / false |
The repository contains full Hugging Face safetensors weights, configuration, tokenizer, processor configuration, and chat template. It is not a LoRA adapter. Load using a Transformers version that supports qwen3_5 / Qwen3_5ForConditionalGeneration and preserve the provided chat template.
The model is intended to produce thought-and-action text for interaction with ScienceWorld. It does not include the ScienceWorld environment or JAR. Task success must be evaluated in the intended environment; no benchmark success-rate claim is made for this checkpoint in this model card.
Optimizer states, RNG states, trainer state, training logs, local paths, and credentials are not included. These files support model loading but not exact optimizer-state resumption.