litillabs

litil-legal-request-router-1.5b

README

License: apache-2.0

What this model does

LiTiL Legal Request Router turns an incoming legal request into a consistent application route. It classifies the request as commercial, employment, corporate, technology/AI, crypto, or general legal work and returns the choice as JSON that a workflow can validate directly.

Place it at the front of a legal-intelligence stack. The accepted route can select the right intake form, retrieval index, playbook, specialist model, or review queue before expensive analysis begins. Storing the route with the original request also makes volume, reassignment, and fallback behavior measurable across the intake system.

Useful for: sorting requests before a specialist workflow takes over
Give it: the request text
It returns: a JSON route for commercial, employment, corporate, technology/AI, crypto, or general legal work

Table with columns: At a glance, Detail
At a glance	Detail
Release repository	`litillabs/litil-legal-request-router-1.5b`
Model format	Rank-16 PEFT LoRA adapter, selected checkpoint 200
Base	`Qwen/Qwen2.5-1.5B-Instruct`
Input → output	System instruction and request → one JSON routing proposal
Decoding	Greedy, up to 256 new tokens
Tested runtime	Apple MPS float32 · PyTorch 2.12.0 · Transformers 5.9.0 · PEFT 0.19.1

Input and output contract

Input is a normal Qwen chat messages list using the exact system instruction above. The output must be one unfenced JSON object with these keys:

Table with columns: Key, Type, Accepted values
Key	Type	Accepted values
`domain`	string	Application domain code
`subdomain`	string	Application subdomain code
`complexity`	string	`simple`, `medium`, or `complex`
`tools`

Validate the JSON, required keys and known domain codes before dispatch. Treat confidence, tools and escalate as proposals that the application may replace with its own policy.

Recorded example

json
{
  "domain": "employment",
  "subdomain": "handbook_update",
  "complexity": "medium",
  "tools": [
    "policy_comparator",
    "jurisdiction_checker",
    "termination_risk_assessor"
  ],
  "confidence": 0.75,
  "escalate": false,
  "reasoning": "Domain: employment (handbook_update). Complexity: medium. Escalate=False per v3 rules; reasons=[]."
}

This was produced by the selected adapter on September 10, 2026.

Practical use

Put one request in the user message and keep the system instruction unchanged.
Use greedy decoding and parse the whole completion as JSON.
Reject malformed objects and unknown domain codes.
Map the accepted domain to an application route.
Apply the application's current escalation and tool-selection policy.

Training data and method

The selected checkpoint comes from three-epoch supervised LoRA tuning on 540 authored synthetic examples, with 60 synthetic validation rows. No private client or user data was used in the reviewed post-training set.

LoRA uses rank 16, alpha 32, and dropout 0.05. The full run recorded 204 steps; this release selects checkpoint 200.

Runtime and version

The adapter is approximately 17.5 MB and requires the Qwen2.5-1.5B-Instruct base. The prepared runner used float32 for reproducibility; a compatible lower precision can reduce memory after separate validation.

Base repository: Qwen/Qwen2.5-1.5B-Instruct
Tested base snapshot: 989aa7980e4cf806f80c7fef2b1adb7bc71aa306
Card version: September 11, 2026
Organization: LiTiL Labs

Use the model

bash
python -m pip install "torch==2.12.0" "transformers==5.9.0" "peft==0.19.1"

python
import json
import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

BASE_ID = "Qwen/Qwen2.5-1.5B-Instruct"
BASE_REVISION = "989aa7980e4cf806f80c7fef2b1adb7bc71aa306"
ADAPTER_ID = "litillabs/litil-legal-request-router-1.5b"

messages = [
    {
        "role": "system",
        "content": (
            "You are a legal query router. Analyze the user's legal query and "
            "output a JSON object with: domain, subdomain, complexity "
            "(simple/medium/complex), tools (list), confidence (0-1), "
            "escalate (boolean), reasoning (string)."
        ),
    },
    {
        "role": "user",
        "content": "Review an employee leave policy and summarize the changes proposed by HR.",
    },
]

tokenizer = AutoTokenizer.from_pretrained(BASE_ID, revision=BASE_REVISION)
base = AutoModelForCausalLM.from_pretrained(BASE_ID, revision=BASE_REVISION)
model = PeftModel.from_pretrained(base, ADAPTER_ID).eval()
inputs = tokenizer.apply_chat_template(
    messages, tokenize=True, add_generation_prompt=True, return_tensors="pt"
)

with torch.inference_mode():
    tokens = model.generate(
        inputs,
        do_sample=False,
        max_new_tokens=256,
        pad_token_id=tokenizer.eos_token_id,
    )

text = tokenizer.decode(tokens[0, inputs.shape[-1]:], skip_special_tokens=True).strip()
route = json.loads(text)
print(route)

For the prepared offline package:

bash
python demo.py --show-recorded --case employee_leave
python demo.py --base ./Qwen2.5-1.5B-Instruct --adapter . --case employee_leave

The runner supports cpu, mps and cuda and never executes a proposed tool.

Citation

When using this adapter, cite the release repository and Qwen2.5-1.5B-Instruct.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

litillabs

Model Tree

Base

Qwen/Qwen2.5-1.5B-Instruct

Adapter

this model

Input Modalities

Text

Output Modalities