Discerna-v1-475M

Discerna is a small decision model for text and images. You give it a state (any text or JSON), optional images, and named questions. It answers every question with a probability for each option, in one forward pass, without generating text.

• Three kinds of question: yes/no (noul), one of 2–255 named options (choice), and a level on a scale of 2–10 (score).
• Text and images together: up to 16,384 tokens of text. Each image is read in up to ten 512-pixel tiles plus a thumbnail; requests with images keep up to 896 tokens of text.
• Calibrated probabilities, one temperature per question type and option count.
• Many questions about the same state in one request.
• 475M parameters. Runs on CPU, Apple Silicon and NVIDIA GPUs; the weights take about 1 GB in half precision (the default on GPUs) and 1.9 GB in full precision (the default on CPU).

This repository holds the model and the discerna package that runs it (src/), with examples.

Quick start

Install the package (Python 3.12 or later):

pip install https://huggingface.co/CrazyH0rse/Discerna-v1-475M/resolve/main/wheels/discerna-1.0.0-py3-none-any.whl

Start the server. The first start downloads the model (1.9 GB); later starts use the cached copy:

discerna serve --model CrazyH0rse/Discerna-v1-475M

Get the examples (requests, photos and an answer key) and ask the server from another terminal:

hf download CrazyH0rse/Discerna-v1-475M --include "examples/*" --local-dir .
discerna ask examples/requests/support_ticket.json
discerna ask examples/requests/market_potatoes.json --image examples/images/market_potatoes.jpg --answers examples/answers.json

--image attaches any picture, and --answers marks each answer right or wrong from an answer key. To answer one request without a server:

discerna decide --model CrazyH0rse/Discerna-v1-475M examples/requests/support_ticket.json

--model also takes a local folder, such as a full copy of this repository (hf download CrazyH0rse/Discerna-v1-475M --local-dir Discerna-v1-475M).

Python

import discerna

model = discerna.load("CrazyH0rse/Discerna-v1-475M")  # picks the device and precision
response = model.decide({
    "state": {"channel": "email", "message": "I was charged twice for my March invoice. Please fix this before Friday."},
    "questions": {
        "team": {"type": "choice", "instructions": "Which team should handle this?",
                 "criteria": {"billing": "payments, invoices, refunds", "technical": "bugs, errors, outages",
                              "sales": "pricing and new plans"}},
        "urgent": {"type": "noul", "instructions": "Does this need a reply today?"},
        "mood": {"type": "score", "instructions": "How upset is the customer?",
                 "criteria": ["calm", "annoyed", "angry"]},
    },
})
print(response.answers["team"].choice)  # billing
print(response.answers["urgent"].noul)  # 0.358: the probability of "yes"

Images go in "images": [...]: PIL images or file paths (pathlib.Path) in Python, base64 or data URLs over HTTP. model.decide_batch([...]) answers several requests together, which is faster on a GPU.

Requests and answers

The server takes the same JSON at POST /v1/decide.

Field	Content
state	the input: text or any JSON value (optional)
images	pictures, in order (optional)
questions	{name: question}: one or more questions about the same state

Each question has a type, its instructions, and criteria:

• noul: criteria optional, e.g. {"yes": "...", "no": "..."} to describe each side. The answer is noul, the probability of yes.
• choice: criteria are {"option name": "description or null"}. The answer has choice (the most likely option), probabilities for every option, and confidence.
• score: criteria list the levels from lowest to highest. The answer has probabilities per level ("0", "1", …), score (the expected level) and a legend.

The example above answers:

{
 "model": "Discerna-v1-475M",
 "answers": {
  "team": {"type": "choice", "choice": "billing",
           "probabilities": {"billing": 0.998, "technical": 0.0019, "sales": 0.00004}, "confidence": 0.998},
  "urgent": {"type": "noul", "noul": 0.358},
  "mood": {"type": "score", "score": 0.64, "probabilities": {"0": 0.375, "1": 0.608, "2": 0.017},
           "confidence": 0.608, "legend": {"0": "calm", "1": "annoyed", "2": "angry"}}
 },
 "usage": {"input_tokens": 192, "output_tokens": 0}
}

A request beyond the model's limits gets HTTP 413 with "unsupported" in the body; an invalid request gets 422.

How it works

Each question is read together with the state as one sequence: the state, the question, then each option behind a marker. One forward pass scores every marker, and a softmax over the options gives the probabilities.

Part	Parameters	What it does
Text encoder	354.5M	16 bidirectional layers (10 short-convolution, 6 attention), width 1,024
Vision tower	85.8M	SigLIP2; reads each image tile
Projector	8.4M	maps image features into the text encoder; image positions come first
Decision head	26.2M	question-type embedding, 2 transformer layers, a scorer at each option marker

Training

Training ran in five stages (listed under provenance in config.json):

1. text decisions, all weights trained (3,000 steps of 32 examples);
2. more text decisions, adding generated tasks (1,600 steps of 32);
3. vision, through the projector and image-only adapters, with the vision tower frozen (3,000 steps of 16 image examples);
4. text and images together, the adapters merged and all weights trained (2,500 steps of 32);
5. temperature calibration on held-out text decisions.

Over all stages it trained on about 275,000 examples, counting repeats. Every source permits commercial use; sources under CC BY 4.0 are credited here as their licenses ask.

Source	Publisher	License
MultiNLI (fiction genre left out)	NYU	OANC and per-genre licenses
WANLI	Liu et al.	CC BY 4.0
CommonsenseQA	Talmor et al.	MIT
QASC	Allen Institute for AI	CC BY 4.0
Civil Comments	Jigsaw / Google	CC0 1.0
GoEmotions	Google Research	Apache 2.0
HelpSteer2, HelpSteer3	NVIDIA	CC BY 4.0
Aegis AI Content Safety 2.0	NVIDIA	CC BY 4.0
MASSIVE 1.1	Amazon	CC BY 4.0
PubMedQA (artificial subset)	Jin et al.	MIT
TyDi QA	Google Research	Apache 2.0
ToolACE	Team-ACE	Apache 2.0
Glaive function calling v2	Glaive AI	Apache 2.0
DOCCI	Google LLC	CC BY 4.0
CLEVR v1.0	Johnson et al.	CC BY 4.0

• Generated decision tasks and sentence pairs were written by Qwen/Qwen3-30B-A3B-Instruct-2507 (Apache 2.0) and checked by a second labeling pass.
• Synthetic shape scenes were drawn by code; they contain no third-party data.

Limitations

• It is a small model. It does best on everyday decisions where the answer is in the input: routing messages, yes/no checks, simple ratings, intents.
• Images: it handles clear, simple scenes (a few objects, colors, positions). It is often wrong on busy real-world photos, small or handwritten text, counts of many objects and fine details, and it can be confidently wrong.
• Request wording: it works best with meaningful option names and the text in state. With options named only by letters or numbers, and the text inside the question, accuracy drops.
• Unsure answers: when unsure, it leans toward cautious options such as "none of these" or "neutral".
• Language: most of its training data is English.
• Probabilities: they are calibrated on general data. Check them on your own examples before relying on them.
• Stakes: do not use it for high-stakes decisions without human review.

License

• The model (model.safetensors, with config.json and the tokenizer files) and the model code in src/discerna/model/ are under the LFM Open License v1.0. Commercial use is licensed only to organizations below 10 million US dollars in annual revenue.
• All other code is under the Apache License 2.0.

Discerna is trained further from Liquid AI's LFM2.5-Encoder-350M (text encoder and tokenizer) and LFM2.5-VL-450M (vision encoder and projector). Its model code is adapted from Liquid AI's open d1-omni-600M implementation. NOTICE credits the base models, the training data and the example images.