Instructions to use mainger/mainger-qwen2.5-1.5b-ft with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use mainger/mainger-qwen2.5-1.5b-ft with Transformers:

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

pipe = pipeline("text-generation", model="mainger/mainger-qwen2.5-1.5b-ft")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("mainger/mainger-qwen2.5-1.5b-ft")
model = AutoModelForCausalLM.from_pretrained("mainger/mainger-qwen2.5-1.5b-ft")
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

vLLM

How to use mainger/mainger-qwen2.5-1.5b-ft with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "mainger/mainger-qwen2.5-1.5b-ft"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "mainger/mainger-qwen2.5-1.5b-ft",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/mainger/mainger-qwen2.5-1.5b-ft

SGLang

How to use mainger/mainger-qwen2.5-1.5b-ft 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 "mainger/mainger-qwen2.5-1.5b-ft" \
    --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": "mainger/mainger-qwen2.5-1.5b-ft",
		"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 "mainger/mainger-qwen2.5-1.5b-ft" \
        --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": "mainger/mainger-qwen2.5-1.5b-ft",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use mainger/mainger-qwen2.5-1.5b-ft with Docker Model Runner:
```
docker model run hf.co/mainger/mainger-qwen2.5-1.5b-ft
```

mainger-qwen2.5-1.5b-ft

Fine-tuned Qwen2.5-1.5B-Instruct for the mainger-agent — an LLM agent that drives the mainger R package for Mahalanobis-integrated regression under privacy constraints.

This is an anonymous release accompanying a submission to STAIX 2026.

Intended use

Tool-calling assistant for the mainger-agent: parses user requests about external-data integration, selects a sharing regime (full / partial / restricted), calls into the R package via the agent's tool layer, and explains the returned diagnostics. Not intended as a general-purpose chat model.

How to load

The model was fine-tuned with LoRA and the adapter has been merged into the base weights, so it loads exactly like any Qwen2.5-1.5B checkpoint:

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "mainger/mainger-qwen2.5-1.5b-ft",
    torch_dtype="bfloat16",
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("mainger/mainger-qwen2.5-1.5b-ft")

Serving with vLLM (extended 65k context)

The agent uses YaRN rope-scaling to extend the native 32k context to 65k. Tool calling uses the hermes parser:

vllm serve mainger/mainger-qwen2.5-1.5b-ft \
  --served-model-name qwen-1.5b-ft \
  --max-model-len 65536 \
  --rope-scaling '{"rope_type":"yarn","factor":2.0,"original_max_position_embeddings":32768}' \
  --enable-auto-tool-choice \
  --tool-call-parser hermes \
  --dtype bfloat16 \
  --gpu-memory-utilization 0.85

Note on the rope-scaling key. The JSON uses "rope_type" (not "type"). HF transformers ≥ 4.45 renamed the field; vLLM ≥ 0.6.6 follows the new convention. Older vLLM or transformers may still accept "type"; on a mismatched pair you'll see KeyError: 'rope_type' at startup.

Compatible toolchain pin

vLLM 0.6.6.post1 expects transformers == 4.47.x. With a newer transformers the tokenizer call all_special_tokens_extended is missing and vLLM crashes on startup. Pin the matching pair:

pip install "vllm>=0.6,<0.7" "transformers==4.47.1"

Training

Base model: Qwen/Qwen2.5-1.5B-Instruct
Method: LoRA fine-tuning, merged back into base weights
Precision: BF16
Task: Tool-calling for the mainger-agent's R-package interface

Reproducibility

The agent that uses this model fixes cv_seed = 548 when calling mainger::mainger(..., tuning = "cv", cv_seed = 548), producing bit-identical integration results across machines.

License

Apache 2.0 (inherited from Qwen2.5-1.5B-Instruct).

Downloads last month: 38

Safetensors

Model size

2B params

Tensor type

BF16

Model tree for mainger/mainger-qwen2.5-1.5b-ft

Base model

Qwen/Qwen2.5-1.5B

Finetuned

Qwen/Qwen2.5-1.5B-Instruct

Finetuned

(1594)

this model