barga

barga (Nepali बर्ग, "category") is a small system-one decision model for English and Nepali. You give it a state (a conversation, a document, a game position, a record) and one or more questions, each with its own candidate options. It returns a probability for every option, in one fast forward pass. It does not generate text.

This package is the inference code. The weights are on the Hugging Face Hub at ampixa/barga, and the model card has the evaluation and limitations.

Install

pip install barga               # PyTorch backend
pip install "barga[onnx]"       # optional onnxruntime backend for exported ONNX bundles

Requires Python ≥ 3.10, PyTorch ≥ 2.6 and transformers 5.17 or 5.18. Those transformers versions are the ones verified to reproduce the reference predictions exactly; see "Verifying a bundle" below.

Quickstart

import barga

model = barga.load("ampixa/barga")          # Hub repo id, or a local bundle directory; device="cuda" for a GPU

state = ("Caller: Hello, I want to book a check-up for my dog tomorrow morning.\n"
         "Agent: Sure. We have 9:30 or 11:00 tomorrow.\n"
         "Caller: 9:30 is fine.")

decisions = model.decide(state, [
    {"id": "intent", "text": "What does the caller want?",
     "options": [{"id": "book", "description": "Book an appointment"},
                 {"id": "cancel", "description": "Cancel an appointment"},
                 {"id": "info", "description": "Only ask for information"}]},
    {"id": "slot", "text": "Which time did the caller choose?",
     "options": [{"id": "t0930", "description": "9:30 tomorrow"},
                 {"id": "t1100", "description": "11:00 tomorrow"},
                 {"id": "none", "description": "No time chosen yet"}]},
    {"id": "intent_ne", "text": "कल गर्नेले के चाहनुहुन्छ?",
     "options": [{"id": "book", "description": "अपोइन्टमेन्ट बुक गर्न"},
                 {"id": "cancel", "description": "अपोइन्टमेन्ट रद्द गर्न"},
                 {"id": "info", "description": "जानकारी मात्र सोध्न"}]},
])
for d in decisions:
    print(d.question_id, d.choice, d.probs)

Output from the released weights (probabilities rounded here):

intent    book   {'book': 0.989, 'cancel': 0.009, 'info': 0.003}
slot      t0930  {'t0930': 0.989, 't1100': 0.010, 'none': 0.001}
intent_ne book   {'book': 0.983, 'cancel': 0.002, 'info': 0.015}

The state is encoded once and shared by all of its questions, so asking several questions about the same state costs much less than asking them one at a time.

Questions and options

Each question is a dict:

Option ids are never shown to the model. Only the question text, the rubric and each option's description are. Ordinal questions also return expected_value (the probability-weighted option value).

The model reads at most 1,024 state tokens (when a state is longer, barga keeps its beginning and its most recent end) and 768 tokens for a question with all of its options. A question that does not fit is rejected with barga.SerializationError; options are never dropped or cut.

See docs/RECORDS.md for the full input format and how to score canonical barga records with decide_record.

Reading the output

decide() returns one barga.Decision per question, in order:

• choice: the most probable option id
• probs: option id → probability, in your option order; sums to 1
• expected_value: for ordinal questions

Probabilities compare options within a question. A low top probability (say below 0.6) means the model is unsure, which is often right for questions the state does not answer. Calibrate a threshold on your own data before acting on it automatically.

Verifying a bundle

Every load checks the sha256 of every file against the bundle's bundle.json and refuses a mismatch. The package's parity test reproduces barga's reference evaluation bit for bit on real records:

pip install "barga[test]"
BARGA_TEST_BUNDLE=/path/to/bundle BARGA_TEST_PARITY=/path/to/parity-dir pytest tests/test_bundle_parity.py

Licence

Apache-2.0 (see LICENSE and NOTICE). The weights are fine-tuned from Julia-1 (Supersonic Labs, Apache-2.0), which is built on mmBERT-small (JHU CLSP, MIT).