USAGE.md

Using Numo-1-E2B

Overview · Tasks · Output

For literature organization, evidence screening and research-assistant workflows, see why Numo. Named tasks return defined answers; your workflow retains responsibility for human review and domain validation.

Setup

The model folder contains all inference weights and the tokenizer. It does not need a separate Gemma download, private Numo code or PEFT for inference.

Table as text

Tested software	Version
Python	3.12
PyTorch	2.14.0
Transformers	5.17.0

In a Linux Python environment, install the versions in requirements.txt. Choose the appropriate PyTorch CUDA wheel for your hardware. The tested GPU is an RTX 3060 with 12 GB VRAM, on a 64 GB RAM host. Other versions and hardware have not been verified against this release. A minimum system-RAM configuration has not been measured.

Prerequisite: a C compiler and Python development headers. On the first GPU decision, PyTorch 2.14 compiles a small Triton helper, and that needs gcc and Python.h. Without them, loading works but the first prediction fails with fatal error: Python.h: No such file or directory. Either install them (Ubuntu: sudo apt install build-essential python3-dev) or create your environment from a Python build that ships its headers, such as one installed by uv or conda.

Allow CPU memory for the full 9.27 GB of inference tensors and loading overhead. During GPU inference, a large lookup table stays on CPU; that table alone uses about 4.7 GB. Load the model on CPU first, then call prepare_for_inference("cuda"). The standalone reload used 4.31 GiB peak PyTorch CUDA allocation for nine synthetic requests; this is not total-device VRAM or a maximum-length guarantee. See the README example.

From the model folder:

python example.py
python example.py --device cpu

The second command uses CPU placement. Full-model CPU speed has not been measured. Automatic device maps, FP16, quantization, arbitrary batching and alternative attention backends are not validated release settings.

Choose a task

Call model.predict(tokenizer, task=..., text=...). The model supplies the declared question and answer choices for that task.

Table as text

Task ID	Answers	Recommended text
scientific_field	2–8 selected broad arXiv fields	Title and abstract
research_area	Medicine, biology, physics, computing or another area	Title and abstract
study_approach	Six declared study approaches	Title and abstract
biomedical_relevance	Irrelevant, partially relevant or relevant	Seed and candidate paper
support_boolean	False or true: evidence fully supports the hypothesis	Evidence and hypothesis
support_grade3	None, partial or full support	Evidence and hypothesis
consensus4	Four coded vote-consensus bands	Evidence and hypothesis
consensus5	Five coded vote-consensus bands	Evidence and hypothesis

Use these text formats:

paper = "Title: ...\nAbstract: ..."
evidence = "Evidence:\n...\nHypothesis:\n..."
pair = "Seed paper:\n<title>\n<abstract>\nCandidate paper:\n<title>\n<abstract>"
result = model.predict(tokenizer, task="research_area", text=paper)

The exact descriptions, IDs and meanings are in task-contracts.json and model.config.task_contracts. For scientific_field, pass candidate_ids=[...] to choose or reorder a declared subset. Other calibrated tasks require their fixed answer order.

Read the result

Table as text

Field	Meaning
answer	Description of the selected answer
answer_id	Stable answer ID for a named task
semantic_answer	Its declared meaning, such as a Boolean value
probabilities_by_id	Probability scores keyed by answer ID
calibrated	Whether the exact request matches a calibrated task
review_required	Always true: a person or your workflow decides what to accept
complete_input_tokens	Length of the full rendered request
backbone_calls	One model pass per decision

Probability scores are adjusted using the release's existing calibration. They are not a guarantee of correctness. See evaluation.

Input rules

The full request must fit 1,024 tokens, including instructions and choices. Text, question and choices must be nonempty. There must be 2–8 distinct choices. Oversized input and reserved control tokens are rejected without truncation. Boolean answers use false/true order; ordinal levels use ascending order.

Custom questions

For a question outside the named tasks, supply a typed request:

result = model.predict(tokenizer, request={
    "text": "A synthetic example describes two proposed methods.",
    "question": "Which method is described?",
    "choices": ["Method A", "Method B"],
    "kind": "choice"
})

For a yes/no question use "kind": "noul"; for ordered levels use "kind": "score". Custom requests outside the exact declared definitions return uncalibrated research output. Similar wording alone does not make a request calibrated.

More synthetic requests are in examples.json.