LiquidAI / LFM2.5-Encoder-350M-Policy-Linter

Try LFM • Docs • LEAP • Discord

LFM2.5-Encoder-350-Policy-Linter

A full fine-tune of LFM2.5-Encoder-350M with a rule-matching head that scores every text token against free-text policy rules in a single encoder pass.

Find more details about our encoders in our blog post .

💻 Demos : Try this fine-tuned model running in a CPU-only Hugging Face space: Zero-shot policy linting ** — check text against your company's rules, written as free text. It scores every token against every rule in one pass.

Usage

⚠️ Loads custom code via trust_remote_code=True (the model wraps a trust_remote_code encoder).

Install the required packages:

pip install torch transformers

Run zero-shot policy linting:

import sys
from pathlib import Path

import torch
from transformers import AutoTokenizer

repo = Path(".")
sys.path.insert(0, str(repo))

from train_bizlint_v02 import Lfm2BidirForRuleMatching

rules = [
    "Flag direct mentions of competitor companies.",
    "Flag promises about guaranteed financial returns.",
]
text = "Our product is better than AcmeAI and will guarantee 30% savings."

prefix = "Policy:\n" + "\n".join(f"- {rule}" for rule in rules) + "\n\nText:\n"
full_text = prefix + text

tokenizer = AutoTokenizer.from_pretrained(repo, trust_remote_code=True)
model = Lfm2BidirForRuleMatching.from_pretrained(repo, trust_remote_code=True).eval()

enc = tokenizer(full_text, return_offsets_mapping=True, return_tensors="pt")
offsets = enc.pop("offset_mapping")[0].tolist()

rule_pool = torch.zeros(1, len(rules), len(offsets))
pos = len("Policy:\n")

for rule_idx, rule in enumerate(rules):
    start = pos + 2
    end = start + len(rule)
    token_idxs = [
        i for i, (a, b) in enumerate(offsets)
        if a < end and b > start and a != b
    ]
    rule_pool[0, rule_idx, token_idxs] = 1 / len(token_idxs)
    pos = end + 1

with torch.no_grad():
    probs = model(**enc, rule_pool=rule_pool)["logits"].sigmoid()[0]

text_start = len(prefix)

for token_idx, (a, b) in enumerate(offsets):
    if b <= text_start or a == b:
        continue

    token_text = full_text[a:b]
    for rule_idx, prob in enumerate(probs[token_idx]):
        if prob.item() > 0.5:
            print(f"{token_text!r} -> {prob.item():.3f}: {rules[rule_idx]}")

📬 Contact

• Got questions or want to connect? Join our Discord community
• If you are interested in custom solutions with edge deployment, please contact our sales team .

Citation

@article{liquidAI2026Encoders,
  author = {Liquid AI},
  title = {LFM2.5-Encoders: Fast at Long Context, Even on CPU},
  journal = {Liquid AI Blog},
  year = {2026},
  note = {www.liquid.ai/blog/lfm2-5-encoders},
}