d3-mini for 🤗 Transformers

The 255-way answer-code readout of vllm-sr/d3-mini as a JEV decision head, for the decision pipeline. This repo holds the readout, the tokenizer and the prompt; the backbone loads from the original repo, with bidirectional full attention as it was trained.

Questions

Each question is answered with a probability for every option: noul (yes/no), choice and score.

from transformers import pipeline

decide = pipeline("decision", model="ArthurZ/d3-mini")
answers = decide(
    {
        "state": {"customer": "Lena", "message": "I was charged twice for my order and I want my money back today."},
        "questions": {
            "refund": {"type": "noul", "instructions": "Does the customer ask for a refund?"},
            "tone": {"type": "choice", "instructions": "What is the tone?", "criteria": {"calm": "Polite", "angry": None}},
            "urgency": {"type": "score", "instructions": "How urgent is it?", "criteria": ["Can wait", "This week", "Today", "Now"]},
        },
    }
)
answers["refund"]["noul"]  # probability of true
answers["tone"]["choice"], answers["tone"]["probabilities"]
answers["urgency"]["score"]  # expected level, 0 to 3

Same questions, new states

A bare state is asked the questions of the last request, or the ones given up front:

decide = pipeline(
    "decision",
    model="ArthurZ/d3-mini",
    questions={"spam": {"type": "noul", "instructions": "Is this spam?"}},
)
decide("WIN A FREE IPHONE!!! click here to claim your prize")
decide("Hi! The parcel arrived, thanks a lot.")

Images

Images are assets, as in the System One API: a placeholder in the state is replaced with its image. An asset is a PIL image, a path, a URL or {"type": "image", "format": "png", "data_base64": ...}.

decide(
    {
        "state": "The customer sent photo-1 and asks if the product is damaged.",
        "assets": {"photo-1": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/coco_sample.png"},
        "questions": {"damaged": {"type": "noul", "instructions": "Is the product in the photo damaged?"}},
    }
)

Converted with convert_decider_to_hf.py. Its prompts are token-for-token those of the original code (decision_test.json), except that control tokens written in a request are escaped instead of read as such.