lfm-2.5-350m-decs1

A System One decision model: instead of generating text, it reads a state and scores every option of every typed question in a single forward pass, returning calibrated probabilities your software can act on. Compatible with the Jev / SystemOne API shape (POST /v1/systemone).

• Backbone: LiquidAI/LFM2.5-350M (350M, hybrid conv-attention, 4-bit QLoRA + fresh Clef-style decision head)
• Trained with: Unsloth FastDecisionModel (DecisionTrainer, LoRA r=16, lr 2e-4, cosine, 3 epochs) on Kaggle 2×T4
• Data: 12,000-row mixture of 13 sources (ag_news, arc, banking77, boolq, clinc150, commonsense_qa, mmlu, mnli, prompt_injections, snli, sst5, wanli, typed-decisions), decontaminated against the test sets

Benchmarks

Accuracy before → after fine-tuning (same protocol as the Unsloth decision guide; test: 2,000 typed-decisions + 500 BANKING77 + 500 CLINC150):

Model	typed-decisions	BANKING77	CLINC150	Holdout acc
lfm-2.5-350m-decs1 (this)	31% → 57%	2% → 68%	17% → 68%	72%
Unsloth Qwen3.5-0.8B (reference)	36% → 73%	7% → 74%	19% → 76%	78%

Calibration on holdout: ECE 0.083. Act on probabilities, not confidence (see below).

Use

A. Unsloth (recommended)

from unsloth import FastDecisionModel

model, tokenizer = FastDecisionModel.from_pretrained(
    "itsdevice/lfm-2.5-350m-decs1", load_in_4bit=True)
FastDecisionModel.for_inference(model)
answers = FastDecisionModel.predict(
    model, tokenizer,
    "Hi, I was charged twice for invoice #4411. Please refund the duplicate today.",
    {"team": {"type": "choice",
              "instructions": "Which team should handle this?",
              "criteria": {"billing": "invoices, payments, refunds",
                           "technical": "bugs, outages, errors",
                           "sales": "pricing, new plans"}},
     "refund": {"type": "noul", "instructions": "Does the customer ask for a refund?"}},
)
print(answers["team"]["answer"], answers["team"]["probabilities"])

B. Decision API (Unsloth Studio / any Jev-compatible server)

curl http://localhost:8888/v1/systemone \\
  -H "Authorization: Bearer sk-unsloth-YOUR_KEY" \\
  -H "Content-Type: application/json" \\
  -d '{"model": "laya",
       "state": "Hi, I was charged twice for invoice #4411. Please refund.",
       "questions": {
         "team": {"type": "choice", "instructions": "Which team?",
                  "criteria": {"billing": "invoices, refunds",
                               "technical": "bugs, errors"}},
         "refund": {"type": "noul", "instructions": "Ask refund?"}}}'

Response shape per question type: choice → {choice, confidence, probabilities}; score → {score, confidence, legend, probabilities}; noul → {noul: P(true)}. usage.output_tokens is always 0 — nothing is generated.

Threshold on probabilities (e.g. act only if P ≥ 0.8), not on confidence: confidence rescales differ between Jev and Laya/Clef-style heads.

Files

File	Purpose
model.safetensors, config.json	Backbone (merged 16-bit)
joint_head.safetensors, joint_head_config.json	Decision head (scores every option jointly)
unsloth_decision_config.json	Calibrated temperatures + eval config
tokenizer.*, chat_template.jinja	LFM2.5 tokenizer

Limits: up to 64 questions per request, 255 options per choice, 10 levels per score. Training context 2048 tokens (serving reads up to 16384); long states are cut from the end, schemas never silently truncated.

Limitations

• 350M-class model: strong on intent/tool routing, not suited for knowledge-heavy, math, or code decisions (same guidance as the base model card).
• English-centric evaluation; multilingual quality follows the LFM2.5 backbone.
• Probabilities are calibrated on the training mixture — re-calibrate (FastDecisionModel.calibrate) after further fine-tuning.

License & citation

Weights under LFM Open License v1.0 (inherited from the base model). If you use this model, please cite the base:

@article{liquidai2025lfm2,
  title={LFM2 Technical Report},
  author={Liquid AI},
  year={2025}
}