Lance1573/acrouter-qwen35-08b-router-lora

ACRouter Qwen3.5-0.8B Router LoRA

This repository contains the Qwen3.5-0.8B PEFT/LoRA router used by Agent-as-a-Router for coding-task model selection. It is an adapter, not a standalone full model. Load it on top of Qwen/Qwen3.5-0.8B.

• Project: https://github.com/LanceZPF/agent-as-a-router
• Dataset: https://huggingface.co/datasets/Lance1573/CodeRouterBench
• Paper: https://huggingface.co/papers/2606.22902
• arXiv: https://arxiv.org/abs/2606.22902

Files

• adapter_model.safetensors: LoRA adapter weights.
• adapter_config.json: PEFT adapter configuration with base model set to Qwen/Qwen3.5-0.8B.
• tokenizer.json, tokenizer_config.json, chat_template.jinja: tokenizer assets copied from the training export.
• training_config.json: compact training hyperparameters.
• eval_metrics.json: ID test metrics for this router.

Training Summary

• Base model: Qwen/Qwen3.5-0.8B
• PEFT type: LORA
• LoRA rank: 16
• LoRA alpha: 32
• LoRA dropout: 0.05
• Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Epochs: 3
• Learning rate: 0.0002
• Max sequence length: 1024
• Training samples: 4483
• Backend model choices: claude-sonnet-4-6, claude-opus-4-6, kimi-k2.5, gpt-5.4, MiniMax-M2.7, qwen3.5-plus, glm-5, Qwen3-Max

Evaluation

Evaluated on the CodeRouterBench ID test split (n=2919):

metric	value
Avg performance	0.474415
Oracle performance	0.570049
Oracle gap	0.095634
Routing accuracy	0.361425
rAcc	0.424460
Strong model call rate	0.373073
Perf/cost ratio	364.336998

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model = "Qwen/Qwen3.5-0.8B"
adapter_id = "Lance1573/acrouter-qwen35-08b-router-lora"

tokenizer = AutoTokenizer.from_pretrained(adapter_id, trust_remote_code=True)
base = AutoModelForCausalLM.from_pretrained(
    base_model,
    torch_dtype="auto",
    device_map="auto",
    trust_remote_code=True,
)
model = PeftModel.from_pretrained(base, adapter_id)
model.eval()

The adapter is intended for model-routing prompts from Agent-as-a-Router rather than general-purpose instruction following.

Limitations

This is a task-specific router trained for selecting among the backend models listed above. It should not be interpreted as a general coding assistant. The adapter does not include private API keys, raw trajectories, optimizer states, or training checkpoints.

Citation

@article{agent2026zhou,
  title         = {Agent-as-a-Router: Agentic Model Routing for Coding Tasks},
  author        = {Pengfei Zhou, Zhiwei Tang, Yixing Ma, Jiasheng Tang, Yizeng Han, Zhenglin Wan, Fanqing Meng, Wei Wang, Bohan Zhuang, Wangbo Zhao, Yang You},
  journal       = {arXiv preprint arXiv:2606.22902},
  year          = {2026},
  archivePrefix = {arXiv},
  eprint        = {2606.22902},
  url           = {https://arxiv.org/abs/2606.22902},
}