You can deploy this model and start earning money today!

franckverrot/
stktc-lfm-adapters

Name: stktc-lfm-adapters
Availability: InStock

LiquidAI/LFM2.5-350M-MLX-4bit

adapter

mit

Views

No views yet

SendThisKidToCollege — LFM2.5-350M LoRA Adapter

LoRA adapter for LiquidAI/LFM2.5-350M-MLX-4bit fine-tuned to parse natural language college search queries into structured JSON filters.

What it does

Takes a query like "bio + D1 soccer under 40k" and produces:

{"reasoning": "bio + D1 soccer + tuition", "majors": ["Biology"], "sport": "Soccer", "division": "NCAA Division I", "tuition_max": 40000}

Performance

Metric	Base LFM2.5-350M	With this adapter
Precision	0.0%	92.8%
Recall	0.0%	92.8%
Value Accuracy	0.0%	93.9%
Parse Errors	30/30	0/30

Evaluated on 30 test cases covering sports, majors, states, divisions, tuition ranges, application systems, and name searches.

Usage with mlx-lm (Python)

python

from mlx_lm import load, generate

model, tokenizer = load(
    "LiquidAI/LFM2.5-350M-MLX-4bit",
    adapter_path="franckverrot/stktc-lfm-adapters"
)

prompt = """You convert a university search query into a JSON filter.
Output ONLY the JSON object, nothing else.

User: "nursing in texas" /no_think"""

output = generate(model, tokenizer, prompt=prompt, max_tokens=200)
# {"reasoning":"nursing + TX","majors":["Nursing"],"state":["TX"]}

Usage with mlx-swift (iOS)

swift

// Load base model
let container = try await LLMModelFactory.shared.loadContainer(
    configuration: ModelConfiguration(id: "LiquidAI/LFM2.5-350M-MLX-4bit")
)

// Download and apply adapter
let hub = HubApi()
let adapterDir = try await hub.snapshot(
    from: Hub.Repo(id: "franckverrot/stktc-lfm-adapters"),
    matching: ["*.safetensors", "*.json"]
)
let adapter = try LoRAContainer.from(directory: adapterDir)
try await container.perform { context in
    try context.model.load(adapter: adapter)
}

Training details

Base model: LiquidAI/LFM2.5-350M-MLX-4bit (4-bit quantized)
Method: LoRA (rank 8, scale 10.0)
Target layers: self_attn.q_proj, self_attn.v_proj (12 layers)
Training: 500 iterations, lr 1e-5, batch size 1
Framework: mlx-lm
Training data: 390 examples of query → JSON filter pairs

Part of SendThisKidToCollege

This adapter is part of the SendThisKidToCollege iOS app — a college search tool that lets families find schools by typing natural language queries like "bio + D1 soccer under 40k."

The app also includes a multi-head DistilBERT classifier (100% precision, 100% recall) that runs via CoreML. This LLM adapter serves as the A/B test alternative.