Instructions to use lugman/LFM2.5-350M-SmolDataAgent with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use lugman/LFM2.5-350M-SmolDataAgent with Transformers:

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

pipe = pipeline("text-generation", model="lugman/LFM2.5-350M-SmolDataAgent")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("lugman/LFM2.5-350M-SmolDataAgent")
model = AutoModelForCausalLM.from_pretrained("lugman/LFM2.5-350M-SmolDataAgent", 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=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use lugman/LFM2.5-350M-SmolDataAgent with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "lugman/LFM2.5-350M-SmolDataAgent"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "lugman/LFM2.5-350M-SmolDataAgent",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/lugman/LFM2.5-350M-SmolDataAgent

SGLang

How to use lugman/LFM2.5-350M-SmolDataAgent 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 "lugman/LFM2.5-350M-SmolDataAgent" \
    --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": "lugman/LFM2.5-350M-SmolDataAgent",
		"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 "lugman/LFM2.5-350M-SmolDataAgent" \
        --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": "lugman/LFM2.5-350M-SmolDataAgent",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Unsloth Desktop
Docker Model Runner
How to use lugman/LFM2.5-350M-SmolDataAgent with Docker Model Runner:
```
docker model run hf.co/lugman/LFM2.5-350M-SmolDataAgent
```

LFM2.5-350M-SmolDataAgent

A 350M-parameter data-analysis agent fine-tuned from LiquidAI/LFM2.5-350M. Given a question and a sandbox of data files (CSV, Parquet, SQLite, HDF5, ...), it works through a single bash tool: explores the files, runs Python, and writes its final answer to /workdir/answer.txt.

Trained specifically on tasks from SmolDataEnvs.

Results

SmolDataEnvs test split: 250 tasks, 4 trials per task, temperature 0.1, up to 12 turns per episode.

Model	pass@1	pass@2	pass@4
LiquidAI/LFM2.5-350M (base)	0%	0%	0%
LFM2.5-350M-SmolDataAgent	21%	26%	30%

By difficulty:

Tier	Tasks	pass@1	pass@2	pass@4
easy	33	56%	61%	64%
medium	118	16%	20%	25%
hard	99	15%	21%	24%

Training data

lugman/SmolDataAgent-Trajectories: 2,833 agent trajectories on SmolDataEnvs train tasks. They started as trajectories from two teacher models, LFM2.5-2.6B and Claude Haiku 5.5, which were then edited (cleaned, and trimmed of tool calls made after the answer was written) and synthetically enriched.

By source and difficulty:

Source	easy	medium	hard	Total
LFM2.5-2.6B	91	1,267	21	1,379
Claude Haiku 5.5	40	140	140	320
Synthetic enrichment	1,007	126	1	1,134
Total	1,138	1,533	162	2,833

By task group:

Source	G1 lookup	G2 multistep	G3 analysis
LFM2.5-2.6B	0	1,362	17
Claude Haiku 5.5	0	0	320
Synthetic enrichment	1,112	19	3

Why not SmolDataEnvs-sft

FineEnvs/SmolDataEnvs-sft is the existing SFT set for this environment. A custom set was generated instead because:

Mixed harnesses. Its rows come from two agent harnesses (bash, seta) with different habits, e.g. 71% of seta rows use a "Plan:" scaffold. Every trajectory here follows one convention.
Lenient grading. Its grader takes the first number anywhere in the answer; 10.4% of re-gradable rows fail the stricter single-value grader used here.
Better results. In a controlled comparison, LFM2.5-350M trained on this custom curated set outperformed the same model trained on SmolDataEnvs-sft.

Training

LoRA (r=16, alpha=32) with Unsloth and TRL, merged into 16-bit weights. Learning rate 4e-4, 3 epochs, effective batch size 16, max sequence length 16,384. Loss on all assistant turns.

Usage

Serve with vLLM and the LFM2 tool-call parser:

vllm serve lugman/LFM2.5-350M-SmolDataAgent --enable-auto-tool-choice --tool-call-parser lfm2

Pass the bash tool in the request and execute its calls in a sandbox. Each call is a fresh, non-stateful bash -c. The episode ends when the model replies without a tool call; read the answer from /workdir/answer.txt.

Recommended sampling: temperature 0.1, top_k 50, repetition_penalty 1.05, max_tokens 1024 per turn.