Token classifier demo

demo_academic_hedged.py

import json

import torch
from huggingface_hub import hf_hub_download
from transformers import AutoModel, AutoTokenizer


def main():
    model_name = "HavelockAI/bert-token-classifier"

    # Load tokenizer + model
    tokenizer = AutoTokenizer.from_pretrained(model_name)
    model = AutoModel.from_pretrained(model_name, trust_remote_code=True)
    model.eval()

    # Load type map
    type_map_path = hf_hub_download(model_name, "type_to_idx.json")
    with open(type_map_path) as f:
        type_to_idx = json.load(f)
    idx_to_type = {v: k for k, v in type_to_idx.items()}

    # The three academic paragraphs
    text = (
        "If one were to assume that the manuscript under consideration, "
        "which might plausibly be interpreted as reflecting certain late-structuralist "
        "commitments, were situated within a broader interdisciplinary discourse, "
        "then it could be argued that its methodological ambiguities would likely "
        "illuminate how epistemic hedging functions across overlapping analytical frameworks.\n\n"
        "In particular, if the reader were to encounter a claim that, insofar as it "
        "may tentatively be understood as provisionally substantiated, appears to "
        "challenge prevailing orthodoxies, then that claim would seem to exemplify "
        "the manner in which layered qualifications can simultaneously signal caution, "
        "reflexivity, and theoretical openness.\n\n"
        "Thus, if it were the case that the argument, which might reasonably be said "
        "to be only partially determinate, were evaluated according to conventional "
        "standards of evidentiary rigor, one might conclude that its apparent "
        "indeterminacy would, under certain interpretive conditions, constitute not "
        "a weakness but a productive site of conceptual negotiation."
    )

    inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=512)

    with torch.no_grad():
        logits = model(**inputs)
        # Use Viterbi decoding if CRF is available, else argmax
        if hasattr(model, "decode"):
            preds = model.decode(inputs["input_ids"], inputs["attention_mask"])
        else:
            preds = logits.argmax(dim=-1)  # (1, seq, num_types)

    tokens = tokenizer.convert_ids_to_tokens(inputs["input_ids"][0])

    print(f"Logits shape: {tuple(logits.shape)}")
    print()

    for i, token in enumerate(tokens):
        active = [
            f"{idx_to_type[t]}={'OBI'[v]}"
            for t, v in enumerate(preds[0, i].tolist())
            if v > 0
        ]
        if active:
            print(f"{token:15} {', '.join(active)}")

    # Simple runtime checks (not pytest)
    assert logits.shape == (
        1,
        inputs["input_ids"].shape[1],
        len(type_to_idx),
        3,
    ), "Unexpected logits shape"

    assert (preds > 0).any(), "Model predicted all-O for every type"


if __name__ == "__main__":
    main()

demo_technical_podcast.py

import json

import torch
from huggingface_hub import hf_hub_download
from transformers import AutoModel, AutoTokenizer


def main():
    model_name = "HavelockAI/bert-token-classifier"

    # Load tokenizer + model
    tokenizer = AutoTokenizer.from_pretrained(model_name)
    model = AutoModel.from_pretrained(model_name, trust_remote_code=True)
    model.eval()

    # Load type map
    type_map_path = hf_hub_download(model_name, "type_to_idx.json")
    with open(type_map_path) as f:
        type_to_idx = json.load(f)
    idx_to_type = {v: k for k, v in type_to_idx.items()}

    # The three podcast-style paragraphs
    text = (
        "So imagine you're reading this manuscript and you're thinking, okay, "
        "this definitely has some late-structuralist vibes going on. If you zoom out "
        "and place it in a bigger interdisciplinary conversation, you start to see "
        "that all those methodological gray areas aren't just accidental. They’re "
        "actually showing you how people hedge their claims differently depending "
        "on which analytical lens they're using.\n\n"
        "Now picture a reader coming across a claim that kind of pushes against "
        "the usual orthodoxy. It’s framed cautiously — lots of 'maybe,' 'arguably,' "
        "and 'to some extent.' That layering of qualifications isn’t just academic "
        "nervousness. It’s doing work. It signals caution, sure, but it also shows "
        "self-awareness and leaves the door open for alternative interpretations.\n\n"
        "And here’s the interesting part: if you judged the whole argument by super "
        "strict, traditional standards of evidence, you might initially think the "
        "fuzziness is a flaw. But under the right interpretive lens, that "
        "indeterminacy becomes productive. It turns into a space where meaning is "
        "negotiated rather than declared — and that’s kind of the point."
    )

    inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=512)

    with torch.no_grad():
        logits = model(**inputs)
        # Use Viterbi decoding if CRF is available, else argmax
        if hasattr(model, "decode"):
            preds = model.decode(inputs["input_ids"], inputs["attention_mask"])
        else:
            preds = logits.argmax(dim=-1)  # (1, seq, num_types)

    tokens = tokenizer.convert_ids_to_tokens(inputs["input_ids"][0])

    print(f"Logits shape: {tuple(logits.shape)}")
    print()

    for i, token in enumerate(tokens):
        active = [
            f"{idx_to_type[t]}={'OBI'[v]}"
            for t, v in enumerate(preds[0, i].tolist())
            if v > 0
        ]
        if active:
            print(f"{token:15} {', '.join(active)}")

    # Simple runtime checks (not pytest)
    assert logits.shape == (
        1,
        inputs["input_ids"].shape[1],
        len(type_to_idx),
        3,
    ), "Unexpected logits shape"

    assert (preds > 0).any(), "Model predicted all-O for every type"


if __name__ == "__main__":
    main()