interfaze-ai/lev

lev answers typed questions about a piece of context in a single forward pass. You give it a state (text, a ticket, an email, or JSON) and a set of yes/no, choice, and score questions. It reads each answer from the logits it already computed and returns calibrated probabilities over exactly the options you supplied. It is a LoRA adapter on Qwen3.5-4B, and it speaks TypeSafe's /v1/systemone protocol, so code written for the TypeSafe SDK works against it once you change the base URL.

74.3% on all 13 S1Bench subsets. 4B parameters. Zero output tokens.

Qwen3.5-4B + LoRA on one H100. On the six subsets the public S1Bench board completed, behind only Jev and two open models of 26B and up, ahead of reflex-4b and a 35B model.

Quickstart · Self-hosting · Benchmarks · Speed · Why it works · Training · Boundaries · GitHub

No generated tokens, no JSON to parse, no retries. lev cannot return a label outside your options, because the answer space is the option set you send. It can still pick the wrong option: the guarantee is structural, not a guarantee of correctness.

It is built for the high-volume judgement calls inside a product: routing, moderation, intent detection, triage, grading, and checking LLM output.

Installation

pip install "lev[serve] @ git+https://github.com/Abhinavexists/lev#subdirectory=packages/lev"

This needs Python 3.12 or newer and, for real-time use, a CUDA GPU. The serve extra installs torch, transformers, peft, and the HTTP server. The first load downloads the base model (Qwen/Qwen3.5-4B, 9.3 GB of weights) and this release (about 205 MB: the 170 MB adapter, the 15 MB Mode B head and the tokenizer).

Quickstart

import lev

model = lev.load("interfaze-ai/lev")

state = "Hi, I was charged twice for my order #4471 and I want a refund."
questions = {
    "intent": {
        "type": "choice",
        "instructions": "What does the customer want?",
        "criteria": {
            "refund": "wants money back",
            "cancel": "wants to cancel an order",
            "track": "wants to know where an order is",
            "other": "anything else",
        },
    },
    "urgent": {"type": "noul", "instructions": "Does this need a human within the hour?"},
    "frustration": {
        "type": "score",
        "instructions": "How frustrated is the customer?",
        "criteria": ["calm", "mildly annoyed", "annoyed", "angry"],
    },
}

result = model.system_one(state, questions)
print(result.answers["intent"].choice)  # refund
print(result.answers["intent"].probabilities)
# {'refund': 0.84, 'cancel': 0.094, 'track': 0.012, 'other': 0.054}
print(result.answers["urgent"].noul)  # 0.43
print(result.answers["frustration"].score)  # 1.57, between "mildly annoyed" and "annoyed"
print(result.usage.output_tokens)  # 0

These are real outputs from this checkpoint. The answers are objects, and result.model_dump() gives the same JSON the HTTP server returns.

All the questions share one forward pass, so asking three questions costs about the same as asking one. lev.load reads lev_release.json from this repository to find the base model, the prompt format the adapter was trained with, and the strength to apply the adapter at. It also applies the shipped calibration and loads the matching head, so there is nothing to configure.

Because the probabilities are calibrated, you can gate on them. For example, act automatically above 0.9 and send anything lower to a person.

Self-hosting: a Jev-compatible HTTP server

lev serve --checkpoint interfaze-ai/lev --host 0.0.0.0 --port 8000

curl -s localhost:8000/v1/systemone -H 'content-type: application/json' -d '{
  "state": "The package arrived crushed and the screen is cracked.",
  "questions": {"damaged": {"type": "noul", "instructions": "Was the item damaged?"}}
}'

Existing TypeSafe clients work once you point them at the server:

from typesafe_sdk import Choice, Noul, TypeSafeClient

# The first request after startup compiles kernels; allow more than the default 10 s.
client = TypeSafeClient(base_url="http://localhost:8000", api_key="local", timeout=60)
response = client.system_one(
    state={"ticket": "The app crashes every time I open settings."},
    questions={
        "team": Choice(
            instructions="Which team owns this?",
            criteria={"billing": "payments, refunds", "technical": "bugs, crashes", "other": None},
        ),
        "bug": Noul(instructions="Is this a bug report?"),
    },
)
print(response.answers["team"].choice, response.answers["bug"].noul)  # technical 0.92

GET /health reports the loaded checkpoint, whether calibration is active, the routing settings, and requests in flight against max_pending. The server batches every question in a request into one forward pass, and concurrent requests into shared forwards, so clients need not batch for throughput. Past --max-pending requests in flight (default 64) it answers 503 at once rather than queueing; a request queued past 30 s gets 504, and a malformed question gets 422 with the reason.

Benchmarks

S1Bench

lev and TypeSafe Jev ran through the same harness on all 3,880 items S1Bench scores, pinned by Nimble's manifests. The Jev run lands within 0.8 points of TypeSafe's published figure on every subset, so the harness is not the gap. The adapter scale was chosen by leave-one-family-out over these subsets; scored only on families held out of that choice, it estimates 0.728 (FINDINGS §19).

• Noise: at these sizes a per-subset difference needs roughly 5–9 points to be real. lev's leads on aegis2, helpsteer2, multinli and summeval-relevance are all inside that.
• Where Jev is clearly ahead: the minimal-edit pairs (vitaminc −7.5, paws −5.6). The adapter runs at a quarter of its trained strength, which narrowed those gaps from −13.3 and −12.4 and closed summeval-consistency's from −54.1 to −0.6: at full strength it rated most fully faithful summaries one level low.
• Constant baselines: on civil_comments (89% not toxic) and the three 5-level rating subsets, always answering the most common label beats both models.
• Calibration: mean ECE 0.097 for lev, 0.091 for Jev; lev is better calibrated on 6 of 13. summeval-consistency is the outlier, under-confident at 0.303.

Held-out split

A held-out split of the 29 training sources, with no row shared with training:

* Measured on this checkpoint before the serving update that skips split label codes. That update routes choice sets of more than 68 options to label-token readout and re-selects the temperatures. The banking77 and clinc_oos rows come from after the update, which raised banking77 from 0.818. Smaller option sets are routed the same way as before.

Speed

A call is one batched forward pass over every question, so compute stays flat from one question to eight, and a 60-option choice costs the same as a yes/no.

The 69 ms is engine compute for a short request (a three-sentence state), measured inside the container; S1Bench's longer states take more. End to end from a laptop, Jev's hosted API answered in 335–346 ms median and lev on one Modal H100 in 414–654 ms across two runs.

Why it works

• Label-token readout. Each option gets a short code, and the answer is read from the next-token logits over those codes. Codes that would split into two tokens are skipped, so every option stays one token. This handles yes/no, scores, and choice lists up to several hundred options. Choices are read in two option orders and averaged, which cancels position bias. Score levels stay in their trained order unless lev serve runs with --score-order-average, which is off by default.
• Candidate-path head. Past that point, a small learned head matches the state against each option's text, so the number of options has no fixed ceiling.
• Calibration. Temperatures fitted after training are applied at load time. There is one per question type, readout mode, and choice option-count band. They were selected by how well they carry over to task families left out of the fit, not only by how well they fit held-out rows.

The optimizations that mattered

• One batched forward instead of prefill-and-fork: 169 → 69 ms. At this size the forward pass is bound by kernel launches, not arithmetic, so forking the cache saved FLOPs and cost time. One batched forward plus the depthwise-conv kernel cut compute by 59%. ADR-023
• Label-token readout up to the tokenizer's limit: +51 points on MASSIVE's 60 intents. Serving 60 options through label codes instead of the learned head took accuracy from 0.231 to 0.746. ADR-025
• Skipping codes that split: banking77 0.818 → 0.980. Passing over codes that tokenize to two tokens lifts label-token readout from 68 options to several hundred. ADR-028
• The backbone's own prompt format: 0.653 → 0.710 frozen. The untuned instruct model scored 5.7 points higher on the earlier six-subset S1Bench set when questions are dressed in its chat template, so the final run trained in that format. ADR-027

Training

• Data: 200,000 examples from 29 sources built on 26 public Hugging Face datasets. The tasks cover topic, sentiment, and emotion classification; intent detection (banking77, clinc_oos, snips); NLI (SNLI, ANLI, FEVER); paraphrase (MRPC, QQP, PARADE, plus word-swapped hard negatives); multiple-choice QA (RACE, ARC, SciQ, OpenBookQA, CommonsenseQA, StrategyQA); toxicity and safety (ToxiGen, ToxicChat, BeaverTails); and helpfulness (UltraFeedback). Questions are paraphrased, negated, and recast between types, and option sets are shuffled and resized, so the model learns the question format and not one wording.
• Contamination guard: the data build refuses any source that resolves to one of the 13 S1Bench subsets.
• Recipe: LoRA r=32, α=64 on the q/k/v/o attention and MLP projections; 3 epochs, 18,750 steps, batch size 32, learning rate 5e-5, in the base model's chat format, on one H100 in 7.8 hours.

Boundaries worth understanding

• Minimal edits and fine-grained ratings are weak. Inputs that differ by one swapped word or number, and quality ratings over five levels, are where lev is least accurate and can be confidently wrong.
• Calibration is fitted on the training distribution. Temperatures are chosen to transfer across task families. Even so, a task very unlike the training mix may be less well calibrated. Check on your own data before you gate on the probabilities.
• Questions are answered independently. Answers in one request do not condition on each other. Encode a joint decision as one choice, or ask in stages.
• English only.
• Needs a GPU for real-time use. It runs on CPU, but a 4B backbone there takes seconds per call, not milliseconds.

Hosted classification

Text classification in Interfaze runs on a similar system to lev: the model reads the answer from the set of labels you define. The difference is that Interfaze is still token based, so it's a hybrid.

Tokens cost a little speed, but they let one request classify a document while also reading, searching, and extracting from it. Define your labels as an enum in the schema, and the label comes back as a typed field.

Interfaze docs → interfaze.ai/docs

Files

License

The adapter is released under Apache-2.0, the same license as the base model. Some of the training datasets have their own terms, including non-commercial licenses. Review them before commercial use.

Not affiliated with or endorsed by TypeSafe AI.

Apache 2.0 · Interfaze