---
license: apache-2.0
base_model: Qwen/Qwen3.5-2B
library_name: peft
pipeline_tag: text-generation
tags: [lora, qlora, peft, tool-calling, transit, metrollm-bench]
---

# Qwen3.5-2B-metro-v24

QLoRA fine-tune of **Qwen3.5-2B** 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/Qwen3.5-2B-metro-v23`.

## Held-out results (n=238, mean of 3 seeds)

| Metric | 2B base | 2B + v24 | Δ |
|---|---:|---:|---:|
| Tier-1 | 74.17 | **81.20** | +7.03 |
| Composite | 71.90 | **79.69** | +7.79 |

The largest relative PEFT gain on the size curve. At 1.2 GB Q4_K_M it runs on a fanless laptop (~39 tok/s sustained on an M2 Air).

## Contents
- `adapter/` — LoRA adapter (rank 16, α 32; QLoRA 4-bit NF4) + tokenizer + chat template
- `Qwen3.5-2B-metro-v24-Q4_K_M.gguf` — merged GGUF (1.2 GB)
- `training_summary.json`

GGUF: \`llama-server --hf-repo continker/Qwen3.5-2B-metro-v24 --hf-file Qwen3.5-2B-metro-v24-Q4_K_M.gguf\`. The LoRA adapter keys use the \`.language_model.\` prefix; strip it to load onto text-only Qwen3.5-2B. Apache 2.0.

## Links

- [Paper (arXiv:2609.10016)](https://arxiv.org/abs/2609.10016)
- [Paper page (Hugging Face)](https://huggingface.co/papers/2609.10016)
- [Technical report v1.2 (PDF)](https://github.com/continker/metrollm-bench/blob/main/paper.pdf)
- [Benchmark and code](https://github.com/continker/metrollm-bench)
- [Live demo](https://huggingface.co/spaces/remcohendriks/metrollm)
- [All four v24 students (collection)](https://huggingface.co/collections/continker/metrollm-bench-v24-6a35b586a11068e1b1ba3d47)

## Citation

```bibtex
@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}
}
```