Since the outputs are constrained to only a fixed set of options, the model can't select anything outside of those. This however does not guarantee that the output will be correct. It's also common to treat the output token probabilities a confidence scores in this context, but without additional training, those scores likely reflect its confidence in what the next token will be rather than the true probability of the response being the correct answer.

Build your own

We can emulate this behavior by constraining output tokens using an LLM. Here I'm using Qwen/Qwen3-1.7B

import argparse
import json

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "Qwen/Qwen3-1.7B"
options = ["A", "B", "C", "D", "E"]

parser = argparse.ArgumentParser()
parser.add_argument("--input", default="question.json")
args = parser.parse_args()

# load the tokenizer and the model
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)

# the token the model would emit for each option as the first assistant token
option_token_ids = [tokenizer.encode(opt, add_special_tokens=False)[0] for opt in options]

def format_prompt(item):
    prompt = item["question"] + "\n"
    for opt in options:
        prompt += f"{opt}. {item[opt]}\n"
    prompt += "Answer:"
    messages = [
        {"role": "user", "content": prompt}
    ]
    return tokenizer.apply_chat_template(
        messages,
        tokenize=False,
        add_generation_prompt=True,
        enable_thinking=False
    )

with open(args.input) as f:
    item = json.load(f)

model_inputs = tokenizer(format_prompt(item), return_tensors="pt").to(model.device)
with torch.no_grad():
    logits = model(**model_inputs).logits[0, -1]
# constrained decoding: only the option tokens are allowed
probs = torch.softmax(logits[option_token_ids].float(), dim=-1)

print(f"prediction: {options[probs.argmax().item()]}")
for opt, prob in zip(options, probs.tolist()):
    print(f"{opt}: {prob:.4f}  {item[opt]}")

Running it with an simple question to test it yeilds the following output

// input

{
    "question": "What color is the sky?",
    "A": "Red",
    "B": "Blue",
    "C": "Green",
    "D": "Purple",
    "E": "I don't know"
}

// output

prediction: B
A: 0.0000  Red
B: 0.9988  Blue
C: 0.0000  Green
D: 0.0000  Purple
E: 0.0012  I don't know

The model was able to make sense of our input and make a prediction that reasonably corresponds to the correct answer.

We can test the accuracy of the model by running it against public datasets. I ran this against a random sample holdout of CommonsenseQA

    precision    recall        f1   support
A    0.5733    0.7197    0.6382       239
B    0.5506    0.7686    0.6416       255
C    0.5372    0.6598    0.5922       241
D    0.7206    0.3904    0.5065       251
E    0.7519    0.4255    0.5435       235
accuracy: 725/1221 = 0.5938
macro f1: 0.5844

Not bad for a 1.7B model, Running a quick finetune on the dataset gives us slightly better performance

    precision    recall        f1   support
A    0.6475    0.6611    0.6542       239
B    0.6113    0.6784    0.6431       255
C    0.6234    0.5975    0.6102       241
D    0.6700    0.5418    0.5991       251
E    0.5808    0.6426    0.6101       235
accuracy: 762/1221 = 0.6241
macro f1: 0.6234

Calibrating your model

Testing the model against a very ambiguous problem demonstrates an interesting problem.

// input
{
    "question": "Where would you most likely find a bat?",
    "A": "Cave",
    "B": "Baseball game",
    "C": "Attic",
    "D": "Zoo",
    "E": "Sporting goods store"
}

// output

prediction: A
A: 0.9978  Cave
B: 0.0004  Baseball game
C: 0.0017  Attic
D: 0.0000  Zoo
E: 0.0001  Sporting goods store

There should be no clear answer here, but treating the output probabilities as a pseudo "confidence" score, shows that the model is extremely overconfident in this answer.

If we bin the confidence score ranges in the eval I ran earlier, we can see that the model's confidence does not match its accuracy. This means that the model is not calibrated.

         bin   count  confidence  accuracy
(0.00, 0.10]       0      0.0000    0.0000
(0.10, 0.20]       0      0.0000    0.0000
(0.20, 0.30]       3      0.2834    0.0000
(0.30, 0.40]      26      0.3761    0.2692
(0.40, 0.50]      41      0.4538    0.2683
(0.50, 0.60]      70      0.5490    0.3286
(0.60, 0.70]      74      0.6476    0.3649
(0.70, 0.80]      77      0.7495    0.4286
(0.80, 0.90]     121      0.8555    0.4711
(0.90, 1.00]     809      0.9855    0.7009

We can notice that the model tends to be extremely overconfident in the 0.9 - 1.0 bin but it's only correct 70% of the time. When it makes a prediction with 0.8 - 0.9 confidence it's only accurate ~40% of the time. This means that the model is generally overconfident in its predictions.

Since our goal is to have the model output scores that is reflective of its accuracy, one method we can use to callibrate it is through temperature scaling. By modifying the temperature value, we can flatten its output probability distribution curve and scale it to approximate its accuracy.

Curve fitting fit the temperature parameter to the model's accuracy, I found 3.797280788421631 as a temp value.

         bin   count  confidence  accuracy
(0.00, 0.10]       0      0.0000    0.0000
(0.10, 0.20]       0      0.0000    0.0000
(0.20, 0.30]      82      0.2712    0.2317
(0.30, 0.40]     217      0.3507    0.3917
(0.40, 0.50]     199      0.4472    0.5126
(0.50, 0.60]     166      0.5475    0.5482
(0.60, 0.70]     139      0.6562    0.5827
(0.70, 0.80]     140      0.7492    0.7714
(0.80, 0.90]     169      0.8507    0.7988
(0.90, 1.00]     109      0.9333    0.9541

This gets us a much better calibration. If you want to play around with this, I made a GitHub repo with scripts that walk you through building a dataset, evaluating, finetuning and calibrating your own model. I encourage pulling it and trying it on other bigger models.