Instructions to use laulauland/Qwen3.5-0.8B-overpass-sft with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use laulauland/Qwen3.5-0.8B-overpass-sft with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="laulauland/Qwen3.5-0.8B-overpass-sft")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("laulauland/Qwen3.5-0.8B-overpass-sft")
model = AutoModelForCausalLM.from_pretrained("laulauland/Qwen3.5-0.8B-overpass-sft", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use laulauland/Qwen3.5-0.8B-overpass-sft with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "laulauland/Qwen3.5-0.8B-overpass-sft"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "laulauland/Qwen3.5-0.8B-overpass-sft",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/laulauland/Qwen3.5-0.8B-overpass-sft

SGLang

How to use laulauland/Qwen3.5-0.8B-overpass-sft with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "laulauland/Qwen3.5-0.8B-overpass-sft" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "laulauland/Qwen3.5-0.8B-overpass-sft",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "laulauland/Qwen3.5-0.8B-overpass-sft" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "laulauland/Qwen3.5-0.8B-overpass-sft",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use laulauland/Qwen3.5-0.8B-overpass-sft with Docker Model Runner:
```
docker model run hf.co/laulauland/Qwen3.5-0.8B-overpass-sft
```

Qwen3.5-0.8B-overpass-sft

SFT warmup of Qwen/Qwen3.5-0.8B for Overpass QL generation from natural language.

This is a merged LoRA adapter (not a raw adapter). Use directly as a base model for further RL training (GRPO).

Training details

Base model: Qwen/Qwen3.5-0.8B
Method: LoRA (rank 32, alpha 64) on q/k/v/o/gate/up/down projections
Steps: 500 (1.3 epochs)
Final loss: 0.575
Token accuracy: 87.2%
Hardware: Apple M-series (MPS), ~5.5 hours

LoRA config

param	value
rank	32
alpha	64
target modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
dropout	0.05
trainable params	~5.8M (0.77% of 753M)

Data

OverpassNL train split, filtered to queries under 300 chars (6,001 examples). Formatted as chat: system prompt + NL input + gold Overpass QL query.

Staniek, M., Fink, N., Mahdavi, S. A., & Riezler, S. (2024). Text-to-OverpassQL: A Natural Language Interface for Complex Geodata Querying of OpenStreetMap. Transactions of the Association for Computational Linguistics (TACL), 12, 1375-1390. arXiv:2308.16060

Base model

Qwen3.5-0.8B: hybrid architecture (Gated Delta Networks + sparse MoE), Apache 2.0.

@misc{qwen3.5,
    title  = {Qwen3.5: Towards Native Multimodal Agents},
    author = {Qwen Team},
    month  = {February},
    year   = {2026},
    url    = {https://qwen.ai/blog?id=qwen3.5}
}

Intended use

Intermediate artifact for the overpass-rl project. SFT teaches Overpass QL syntax so that RL (GRPO) training can start from a viable baseline. Not intended for direct deployment.

Limitations

Only trained on OverpassNL examples; limited tag vocabulary
No multi-turn correction ability (SFT is single-turn)
Compound queries (multiple tags) often incomplete — RL is expected to address this

Downloads last month: 29

Safetensors

Model size

0.8B params

Tensor type

BF16

Model tree for laulauland/Qwen3.5-0.8B-overpass-sft

Base model

Qwen/Qwen3.5-0.8B-Base

Finetuned

Qwen/Qwen3.5-0.8B

Adapter

(183)

this model

Dataset used to train laulauland/Qwen3.5-0.8B-overpass-sft

Paper for laulauland/Qwen3.5-0.8B-overpass-sft

Text-to-OverpassQL: A Natural Language Interface for Complex Geodata Querying of OpenStreetMap

Paper • 2308.16060 • Published Aug 30, 2023