---
license: apache-2.0
language:
- en
base_model: Qwen/Qwen3-1.7B
pipeline_tag: text-generation
library_name: transformers
tags:
- qwen3
- sft
- general-knowledge
- multiple-choice
- cs-552
datasets:
- cais/mmlu
metrics:
- accuracy
---

# General Knowledge Model

This model is a fine-tuned version of [`Qwen/Qwen3-1.7B`](https://huggingface.co/Qwen/Qwen3-1.7B) for the CS-552 Modern NLP course project.

The model targets the **General Knowledge** benchmark, where it answers closed-book multiple-choice factual and reasoning questions. It was trained to return the final answer as a single option letter inside a LaTeX `\boxed{}` expression.

## Intended output format

The model should produce answers in the following format:

```text
\boxed{C}
```

Anything outside `\boxed{}` is treated as reasoning and is not used for scoring by the evaluation pipeline.

## Training procedure

This checkpoint was trained using **Supervised Fine-Tuning (SFT)** with LoRA on top of `Qwen/Qwen3-1.7B`.

The SFT data was formatted as instruction-style multiple-choice examples:

```text
Q: ...
A) ...
B) ...
C) ...
D) ...

Answer: \boxed{C}
```

The current checkpoint was trained on a processed General Knowledge dataset derived from MMLU-style multiple-choice examples.

## Model behavior

The model is optimized for:

- closed-book factual question answering
- multiple-choice reasoning
- final-answer extraction through `\boxed{}`
- concise option-letter responses

The tokenizer chat template was configured with non-thinking mode to encourage concise answers.

## Local validation

On the provided General Knowledge validation snapshot from the course starter repository, this checkpoint achieved:

- Extraction rate: `10/10`
- Accuracy: `6/10`

These validation samples are only a small sanity-check set and are not the hidden evaluation benchmark.

## Framework versions

- Transformers
- PEFT
- PyTorch
- Datasets
- Hugging Face Hub

## Limitations

This is an intermediate SFT baseline, not the final model. It was trained mainly to establish a working General Knowledge pipeline and verify that the model can produce extractable boxed answers. Performance may vary on broader or harder factual reasoning tasks.