GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
arthurcs01
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
arthurcs01
Model Tree
Qwen/Qwen3-4B-Instruct-2507
Input Modalities
Output Modalities
Supported Functionality
Solved episodes on held-out instances.
| Level | What it asks | Base | This adapter |
|---|---|---|---|
| N0 | warm-up, method given | 28/180 | 17/24 |
| N1 | warm-up, method not given | 0/60 | 18/24 |
| N2 | foreign load unit | 0/9 | 0/24 |
| N3 | cracked collar | 0/9 | 2/24 |
| N4 | soaked against dry tests | 0/9 | 15/24 |
python
import torchfrom peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizer base = "Qwen/Qwen3-4B-Instruct-2507"tokenizer = AutoTokenizer.from_pretrained(base)model = AutoModelForCausalLM.from_pretrained(base, dtype=torch.bfloat16, device_map="auto")model = PeftModel.from_pretrained(model, "arthurcs01/geotech-rl-qwen3-4b-sft")The results above came from vLLM, not PEFT. This recipe was checked on the real architecture cut to two layers with random weights, not on the full model.
The tensors are unchanged. Their names were converted to PEFT's format, because prime-rl saves a different one. With the original names PEFT loads the file, warns about missing keys and applies nothing.
Apache 2.0, as the base model. These weights modify Qwen3-4B-Instruct-2507 by the Qwen team.
To cite the work, use the code's DOI: 10.5281/zenodo.23861866. The repository describes how AI was used in the project.
import torchfrom peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizer base = "Qwen/Qwen3-4B-Instruct-2507"tokenizer = AutoTokenizer.from_pretrained(base)model = AutoModelForCausalLM.from_pretrained(base, dtype=torch.bfloat16, device_map="auto")model = PeftModel.from_pretrained(model, "arthurcs01/geotech-rl-qwen3-4b-sft")