decision-0.8b

An open, 0.8B-parameter decision model: a LoRA adapter on Qwen3.5-0.8B-Base, trained in 14 minutes on a free Kaggle T4.

A decision model does not chat. You give it a situation, a question and fixed options. It reads them once and returns a probability for every option: one forward pass, no text generation, no output parsing.

Built by Saksham Arora. Code, training and evaluation: github.com/saksham10arora-dotcom/decision-0.8b.

What it fixes

Stock small models favour whichever option is listed first. One epoch of LoRA, with the option order reshuffled every time an example is seen, mostly removes that bias.

300 held-out questions	Base	Fine-tuned
Accuracy	81.0%	85.3%
Answers that flip when options are reordered	23.0%	8.7%
Yes/no accuracy, "No" listed first	62%	82%
Calibration error (ECE)	0.067	0.028
Probability on the option letters at all	0.44	0.999

Honest read: the accuracy gain is borderline on 300 questions (McNemar p = 0.053). The order-bias fix is the real result. Fine-tuning taught the model how to answer, not new facts.

Use

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = "Qwen/Qwen3.5-0.8B-Base"
tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, dtype=torch.float32)
model = PeftModel.from_pretrained(model, "nerfsaksham/decision-0.8b").eval()

The prompt format (options as A, B, C, D, ending in Answer:) and a ready CLI are in decide.py in the GitHub repo:

python decide.py "Should the agent ask the user before deleting 4,000 emails?" --options Yes No

Training

• LoRA r=16, alpha=32 on the attention and MLP projections, one epoch
• Option order reshuffled on every pass so the answer letter carries no signal
• Training data is not redistributed: prep_data.py in the repo rebuilds it from the original datasets under their own licenses

License

The adapter is a derivative of Qwen3.5-0.8B-Base and is released under Apache 2.0. The training and evaluation code on GitHub is MIT.