WaterSheep

Calibrated decisions for any text. Runs entirely in your browser.

Model · Code · Website · Video · Apache 2.0

Usage

pip install transformers torch

from transformers import pipeline

ws = pipeline(model="samratduttaofficial/WaterSheep", trust_remote_code=True)
ws("I was charged twice.", question="Which team should handle this?",
   options=["billing", "shipping", "support"])

With the Python package, decide returns the answer, its confidence and a probability for every option. ask answers several questions about one text:

pip install git+https://github.com/SamratDuttaOfficial/WaterSheep

from watersheep import WaterSheep

ws = WaterSheep.load("samratduttaofficial/WaterSheep")
ws.ask({
    "state": {"customer": "Priya (premium plan)",
              "message": "Charged twice for order #4411 and the package is 12 days late."},
    "questions": {
        "escalate": {"type": "noul", "instructions": "Should a human agent take over now?"},
        "team": {"type": "choice", "instructions": "Which team should handle this?",
                 "criteria": {"billing": "payments, refunds", "shipping": "delivery problems"}},
        "frustration": {"type": "score", "instructions": "How frustrated is the customer?",
                        "criteria": ["calm", "annoyed", "frustrated", "furious"]},
        "issues": {"type": "multi", "instructions": "Which issues are reported?",
                   "criteria": ["double charge", "late delivery", "damaged item"]},
    },
})

Type	Question	Answer
`noul`	yes/no	probability of yes
`choice`	single choice	the option, with a probability for each
`score`	rating scale	the expected level, with a probability for each
`multi`	multi-label	every option above the threshold, with probabilities

Using Jev?

WaterSheep is an open-source alternative to Jev. Run it as a local server:

pip install git+https://github.com/SamratDuttaOfficial/WaterSheep
watersheep --model samratduttaofficial/WaterSheep --serve

It answers Jev's POST /v1/systemone requests on your machine, and TypeSafe's Python SDK works against it without code changes:

export TYPESAFE_BASE_URL=http://127.0.0.1:8766

Any API key value works locally. Multi-label questions ("type": "multi") work too, as plain JSON. WaterSheep is independent and not affiliated with TypeSafe AI.

Evaluation

Evaluation	Accuracy	ECE
In-distribution test split	77.8%	0.026
Held-out datasets, not seen in training	61.2%	0.043

ECE is the expected calibration error (lower is better).

Reliability diagrams: accuracy against confidence for each question type — Accuracy against confidence for each question type, before (raw) and after calibration.

Error against coverage on the test split and the held-out datasets — Error among the questions answered when the model only answers above a confidence threshold. The dots mark thresholds of 0.70, 0.90 and 0.97.

Benchmarks

Benchmark	Suite	Questions	Accuracy	ECE	In training data
goemotions	sentiment	2,000	22.4%	0.023	other split
hatecheck	safety	2,000	75.1%	0.139	no
legal_abercrombie	legal	95	21.1%	0.316	no
legal_contract_nli_confidentiality_of_agreement	legal	82	69.5%	0.177	no
legal_corporate_lobbying	legal	490	68.4%	0.216	no
legal_cuad_audit_rights	legal	1,216	86.3%	0.041	no
legal_definition_classification	legal	1,337	56.9%	0.279	no
legal_function_of_decision_section	legal	367	24.3%	0.245	no
legal_hearsay	legal	94	56.4%	0.307	no
legal_overruling	legal	2,000	62.5%	0.151	no
legal_personal_jurisdiction	legal	50	50.0%	0.160	no
legal_privacy_policy_qa	legal	2,000	58.9%	0.274	no
legal_proa	legal	95	51.6%	0.379	no
legal_ucc_v_common_law	legal	94	62.8%	0.171	no
prompt_injection	safety	116	91.4%	0.079	other split
xstest	safety	450	73.6%	0.140	no

Training

Base model: answerdotai/ModernBERT-base, fine-tuned with a decision head.
Data: openly licensed public datasets (listed in NOTICE) and synthetic decisions from Qwen3.5-4B.
Calibration: a temperature per question type, fitted on a validation split.

Share of synthetic examples kept after verification, by question type (left) and by family (right).

Limitations

English only.
Long inputs are truncated. The watersheep package and its server mark such answers with "truncated": true.
Rating-scale answers are less accurate than the other types.
Probabilities are calibrated on data like the training data; validate them on your own.
Not for high-stakes decisions (medical, legal, financial, hiring) on its own.

License and citation

Apache 2.0 (LICENSE). Attributions: NOTICE.

@misc{watersheep,
  author = {Samrat Dutta},
  title  = {WaterSheep: calibrated decisions for any text},
  year   = {2026},
  url    = {https://huggingface.co/samratduttaofficial/WaterSheep}
}