Qwen3.5-4B-metro-v24

QLoRA fine-tune of Qwen3.5-4B for the MetroLLM-Bench transit-kiosk task. v24 is the leakage-free retraining used in the MetroLLM-Bench paper (teacher traces drawn only from the 717-case training partition; 238 cases held out). Supersedes continker/Qwen3.5-4B-metro-v23.

Held-out results (n=238, mean of 2 seeds)

The efficiency headline of the paper: at 2.6 GB Q4_K_M, the 4B student exceeds both GPT-5.6 tiers on held-out tier-1 (91.32 vs 90.63 for luna and 90.00 for sol) and ties GPT-5.4 full at maximum reasoning effort (91.37).

Contents

• adapter/ — LoRA adapter (rank 16, α 32; QLoRA 4-bit NF4) + tokenizer + chat template
• Qwen3.5-4B-metro-v24-Q4_K_M.gguf — merged GGUF (2.6 GB)
• training_summary.json

GGUF: `llama-server --hf-repo continker/Qwen3.5-4B-metro-v24 --hf-file Qwen3.5-4B-metro-v24-Q4_K_M.gguf`. The LoRA adapter keys use the `.language_model.` prefix; strip it to load onto text-only Qwen3.5-4B. Apache 2.0.

Links

• Paper (arXiv:2609.10016)
• Paper page (Hugging Face)
• Technical report v1.2 (PDF)
• Benchmark and code
• Live demo
• All four v24 students (collection)

Citation

@techreport{hendriks2026metrollm,
  title       = {MetroLLM-Bench: Evaluating Language Models as Transit Kiosk Runtimes},
  author      = {Hendriks, Remco},
  institution = {Continker},
  type        = {Technical report},
  number      = {v1.2},
  year        = {2026},
  month       = {9},
  doi         = {10.5281/zenodo.21893944},
  eprint      = {2609.10016},
  archiveprefix = {arXiv},
  primaryclass = {cs.LG},
  url         = {https://arxiv.org/abs/2609.10016}
}