GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Anni0401
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Anni0401
Model Tree
Qwen/Qwen3-Embedding-8B
Input Modalities
Output Modalities
Supported Functionality
| Method | PORTS (perplexity-based preference loss + contrastive semantic loss), code adapted from the PORTS authors |
| LoRA | r = 16, alpha = 16, dropout = 0.1, target modules q,k,v,o,gate,up,down_proj, QLoRA enabled |
| Hyperparameters | learning rate 1e-5, lambda 0.3, beta 1, gamma 0.5, 2 epochs, gradient accumulation 32 |
| Data | BFCL train split, 1,000 training samples, 20% held out for validation |
| Frozen inference LLM | Llama-3-8B (provides the preference signal) |
| Checkpoint | end of epoch 2 |
The training scripts and sweep definitions are in the GitHub repository. The PORTS source code belongs to its authors; the repository contains only the files we modified or added, with attribution.
python
from peft import PeftModelfrom transformers import AutoModel, AutoTokenizer base = AutoModel.from_pretrained("Qwen/Qwen3-Embedding-8B", torch_dtype="auto", device_map="auto")tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-Embedding-8B", padding_side="left")model = PeftModel.from_pretrained(base, "Anni0401/qwen3-embedding-8b-ports-lora")With vLLM: vllm serve Qwen/Qwen3-Embedding-8B --enable-lora --max-lora-rank 16 --lora-modules qwen3-embedding-8b-ports-lora=Anni0401/qwen3-embedding-8b-ports-lora.
End-to-end tool-selection results on Wild-Tool-Bench are reported in the GitHub repository together with the evaluation scripts and logs.
Apache-2.0, inherited from the base model. The BFCL data is released under Apache-2.0 by its authors.