---
license: apache-2.0
language:
- en
base_model: Qwen/Qwen3.5-9B
tags:
- transit
- kiosk
- tool-use
- agent
- metrollm-bench
- qwen
- lora
- gguf
- quantized
library_name: peft
pipeline_tag: text-generation
---

> [!WARNING]
> **Superseded by [Qwen3.5-9B-metro-v24](https://huggingface.co/continker/Qwen3.5-9B-metro-v24)** — the leakage-free retraining used in the MetroLLM-Bench paper. This v23 model was trained on the full case set (held-out cases included) and is retained for provenance only.


# Qwen3.5-9B + metro-v23 LoRA

Domain-specialised tool-using agent for transit-kiosk tasks: routing, fare calculation,
disruption advisories, accessibility, multilingual cultural notes, multi-turn context tracking,
and policy adaptation across 6 metro systems (MARTA, BART, CTA, Doha, Taipei MRT, Beijing
Subway).

QLoRA r=16 fine-tune of `Qwen/Qwen3.5-9B` on 790 distilled traces from
Qwen3.5-27B and Qwen3.5-35B-A3B teachers (filtered to tier1 ≥ 90% per case, deduplicated
by case_id, evaluated on the [MetroLLM-Bench](https://github.com/...) v23 harness).

## Files

| File | Purpose |
|---|---|
| `Qwen3.5-9B-metro-v23-Q4_K_M.gguf` (5.3 GB) | Runtime artifact for llama.cpp / Ollama |
| `adapter/` | Raw LoRA adapter (use with PEFT + base Qwen3.5-9B) |
| `training_summary.json` | Hyperparameters, seed, dataset version |

## Eval (v23, 6 systems, Haiku judge for Tier 2)

**Cross-system average**: Tier-1 92.4, Composite 90.0
(+2.2 T1 / +1.4 Comp vs base Qwen3.5-9B)

| System | Tier-1 % |
|---|---:|
| MARTA | 94.0 |
| BART | 90.7 |
| CTA | 93.4 |
| DOHA | 93.1 |
| TAIPEI | 92.6 |
| BEIJING | 90.7 |

## Quickstart (llama.cpp)

```bash
huggingface-cli download continker/Qwen3.5-9B-metro-v23 \
  Qwen3.5-9B-metro-v23-Q4_K_M.gguf --local-dir ./models

llama-server -m ./models/Qwen3.5-9B-metro-v23-Q4_K_M.gguf \
  --port 8080 --ctx-size 32768 --n-gpu-layers 999
```

## Quickstart (PEFT adapter, Python)

```python
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.5-9B", torch_dtype="bfloat16")
model = PeftModel.from_pretrained(base, "continker/Qwen3.5-9B-metro-v23", subfolder="adapter")
tokenizer = AutoTokenizer.from_pretrained("continker/Qwen3.5-9B-metro-v23", subfolder="adapter")
```

## Training

- Base: `Qwen/Qwen3.5-9B`
- Method: QLoRA, rank=16, alpha=32, dropout=0.05
- Targets: q/k/v/o + gate/up/down projections
- Optimizer: AdamW, lr=2e-4, cosine, warmup 5%
- Epochs: 3, effective batch 8 (per_device_train_batch_size=2 × grad_accum=4)
- Max sequence length: 4096
- Seed: 42 (default; multi-seed CI in progress for 27B)
- Dataset: 790 distilled examples, see [continker/metrollm-bench-train-data-v23](https://huggingface.co/datasets/continker/metrollm-bench-train-data-v23)

## Limitations

- Trained on 6 metro systems; generalisation to other systems untested.
- Tool-use schema is specific to the MetroLLM-Bench mock server (route_planner,
  fare_calculator, station_info, disruption_feed, knowledge_base,
  submit_assistant_state).
- Quantised to 4-bit (Q4_K_M); for full-precision behaviour use the adapter on
  bf16 base weights.

## Citation

```
@misc{metrollm-bench-2026,
  title={MetroLLM-Bench: Evaluating LLMs as Prompt-Driven Transit Kiosk Agents},
  author={Hendriks, Remco and contributors},
  year={2026},
  publisher={HuggingFace},
  howpublished={\url{https://huggingface.co/continker}}
}
```