fusellm 0.0.4

Train small LLMs and deploy them for fast structured extraction on CPU

Fuse

Train small LLMs and deploy them for fast structured extraction on CPU.

Fuse lets you pull any GGUF model from HuggingFace, run zero-shot structured extraction with dynamic schemas, fine-tune with LoRA via Unsloth/HuggingFace, and export to GGUF for fast CPU inference. Every extraction comes with per-field confidence scores and optional evidence-span localization. No predefined Pydantic models required.

Install

With uv (recommended)

uv add fusellm

With training support (add reporting for TensorBoard/MLflow metric logging):

uv add "fusellm[training]"
uv add "fusellm[training,reporting]"

Run without installing

# One-shot extraction — no install needed
uvx fusellm extract "Sarah Chen is a 34-year-old architect at Stripe" \
  --model bartowski/Llama-3.2-1B-Instruct-GGUF \
  --fields "name:str,age:int,job_title:str"

# Or with a config file
uvx fusellm extract "SpaceX was founded in 2002" \
  --config extract_company.yaml

With pip

pip install fusellm
pip install "fusellm[training]"

Quick Start

Pull a model from HuggingFace and extract

import fuse

# Auto-downloads the best Q4 GGUF from HuggingFace Hub
backend = fuse.LlamaCppBackend(model_name="bartowski/Llama-3.2-1B-Instruct-GGUF")
extractor = fuse.Extractor(backend)

# Zero-shot structured extraction — no Pydantic model needed
result = extractor.extract_from_fields(
    "Sarah Chen is a 34-year-old software architect at Stripe.",
    {"name": str, "age": int, "job_title": str, "company": str}
)
# result is dict-like (an ExtractionResult):
result.to_dict()    # {'name': 'Sarah Chen', 'age': 34, 'job_title': 'software architect', 'company': 'Stripe'}
result.confidence   # per-field confidence, e.g. {'name': 0.94, 'age': 0.99, ...}

Use a local GGUF model

backend = fuse.LlamaCppBackend(model_path="./models/llama-3.2-1b-q4.gguf")
extractor = fuse.Extractor(backend)

result = extractor.extract_from_fields(
    "John is 30 years old and knows Python and Rust",
    {"name": str, "age": int, "skills": list[str]}
)
result.to_dict()   # {'name': 'John', 'age': 30, 'skills': ['Python', 'Rust']}

Config-driven extraction

config = fuse.InferenceConfig(
    model_name="bartowski/Phi-4-mini-instruct-GGUF",
    n_ctx=4096,
    n_threads=8,
    temperature=0.0,
)
backend = fuse.LlamaCppBackend.from_config(config)

Extract from a JSON schema

schema = fuse.SchemaBuilder.from_json_schema({
    "type": "object",
    "properties": {
        "name": {"type": "string"},
        "age": {"type": "integer"},
        "skills": {"type": "array", "items": {"type": "string"}},
    },
    "required": ["name", "age"],
})
result = extractor.extract("John is 30 and knows Rust", schema)

Let the LLM infer the schema

result = extractor.extract_from_description(
    "The Series A raised $15M from Sequoia, following a $2.5M seed from YC.",
    "Extract monetary amounts, funding round type, and investor names"
)

CLI

Extract with a config file

fuse extract "Sarah Chen is a 34-year-old architect at Stripe" \
  --config examples/extract_person.yaml

extract_person.yaml:

model:
  model_name: "bartowski/Llama-3.2-1B-Instruct-GGUF"
  n_ctx: 2048
  temperature: 0.0

fields:
  name: str
  age: int
  job_title: str
  company: str

prompt_format: llama
max_tokens: 256

Extract with inline flags

# HuggingFace model — auto-downloads
fuse extract "SpaceX was founded in 2002" \
  --model bartowski/Phi-4-mini-instruct-GGUF \
  --fields "company:str,year:int,industry:str"

# Local GGUF model
fuse extract "John is 30" \
  --model ./model.gguf \
  --fields "name:str,age:int"

# Using a JSON schema file
fuse extract "John is 30 and knows Python" \
  --model bartowski/Llama-3.2-1B-Instruct-GGUF \
  --schema schema.json

Train

fuse train --config examples/train_extraction.yaml

Quantize to GGUF

fuse quantize --model ./output --output model.gguf --method q4_0

Supported Models

Any GGUF model on HuggingFace works. Some good small models for CPU extraction:

Model	Size	HuggingFace Repo
Llama 3.2 1B Instruct	~1GB Q4	bartowski/Llama-3.2-1B-Instruct-GGUF
Llama 3.2 3B Instruct	~2GB Q4	bartowski/Llama-3.2-3B-Instruct-GGUF
Gemma 4 E2B Instruct	~3.5GB Q4	bartowski/google_gemma-4-E2B-it-GGUF
Gemma 4 E4B Instruct	~5.4GB Q4	bartowski/google_gemma-4-E4B-it-GGUF
Qwen 2.5 1.5B Instruct	~1GB Q4	bartowski/Qwen2.5-1.5B-Instruct-GGUF
Phi-4 Mini Instruct	~2.5GB Q4	bartowski/Phi-4-mini-instruct-GGUF

Models are auto-downloaded and cached to ~/.cache/fuse/models/.

Development

uv sync --extra dev
uv run nox                    # All CI checks
uv run nox -s lint            # Ruff lint + format
uv run nox -s typecheck       # ty type check
uv run nox -s tests           # Pytest across Python 3.11-3.13

License

MIT