general_knowledge_model/README.md

---
license: apache-2.0
language:
- en
base_model: Qwen/Qwen3-1.7B
pipeline_tag: text-generation
library_name: transformers
tags:
- qwen3
- sft
- dpo
- lora
- general-knowledge
- multiple-choice
- cs-552
datasets:
- cais/mmlu
- TIGER-Lab/MMLU-Pro
- allenai/ai2_arc
- allenai/openbookqa
- allenai/sciq
- tau/commonsense_qa
- allenai/quartz
metrics:
- accuracy
---

# General Knowledge Model

This model is the General Knowledge individual-model submission for the CS-552 Modern NLP course project. It is a merged post-trained checkpoint based on [`Qwen/Qwen3-1.7B`](https://huggingface.co/Qwen/Qwen3-1.7B), developed by Tuan Dang Nguyen for closed-book multiple-choice general knowledge evaluation.

The uploaded checkpoint corresponds to the final Stage 5 merge-aware DPO model:

```text
sft_dpo_stage5_error_contrastive_mergeaware_v1_r16_lr8e8_beta003_eval100_500_merged
```

## Task And Output Format

The model receives a multiple-choice question and should answer with exactly one option letter inside a LaTeX boxed expression:

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

The evaluation pipeline extracts the letter inside `\boxed{...}`. Any surrounding reasoning is ignored for scoring, but the intended behavior is a concise boxed final answer.

## Training Summary

The training campaign used LoRA-based post-training on top of `Qwen/Qwen3-1.7B`.

Main stages:

- Supervised fine-tuning on mixed general-knowledge multiple-choice data.
- Hard-source and CI-style refinements, including MMLU-Pro and variable option-count examples.
- Plus Quartz v1 SFT, which first reached the best hidden-CI score.
- Conservative Stage 2 SFT refinement from the Plus Quartz anchor.
- Stage 5 merge-aware DPO using the Stage 2 model's own wrong boxed answers plus protection pairs.

The final Stage 5 model is selected because it is the strongest merged local checkpoint. The strongest hidden-CI score was first reached by the Plus Quartz SFT anchor, and the later Stage 2/DPO submissions tied that hidden score.

## Evaluation

Local evaluation used the course ten-example public General Knowledge validation snapshot in both prompt modes plus a 290-example diagnostic set built from public multiple-choice sources.

| Model | Role | Local diagnostic | Public 10-example validation | Extraction | Hidden CI |
| --- | --- | ---: | ---: | ---: | ---: |
| `sft_plus_quartz_v1_r128_7200_merged` | First hidden-CI anchor | 247/290 | 7/10 in both prompt modes | 100% | **0.4900** |
| `sft_stage2_plus_quartz_v1_r32_lr5e7_800_merged` | Best retained SFT refinement | 248/290 | 7/10 in both prompt modes | 100% | 0.4900 tie |
| `sft_dpo_stage2_plus_quartz_v1_from_800_mistake_only_r16_lr2e7_beta005_200_merged` | Early DPO refinement | 248/290 | 7/10 in both prompt modes | 100% | 0.4900 tie |
| `sft_dpo_stage5_error_contrastive_mergeaware_v1_r16_lr8e8_beta003_eval100_500_merged` | Uploaded final model | **249/290** | 7/10 in both prompt modes | 100% | 0.4900 tie |

Interpretation: DPO improved the retained merged local diagnostic result and made checkpoint selection more robust, but it did not improve beyond the best hidden-CI SFT score of `0.4900`.

## Usage Notes

This checkpoint is a fully merged model, not a standalone LoRA adapter. It can be loaded with standard `transformers` text-generation tooling.

For best compatibility with the course evaluator:

- Ask closed-book multiple-choice questions.
- Include clear answer options.
- Require the model to finish with `\boxed{LETTER}`.
- Score only the extracted boxed letter.

Example prompt:

```text
Answer the following multiple-choice question. Return only the final answer in the form \boxed{LETTER}.

Question: Which planet is known as the Red Planet?
A) Venus
B) Mars
C) Jupiter
D) Mercury
```

Expected style:

```text
\boxed{B}
```

## Limitations

This model is specialized for English closed-book multiple-choice general knowledge. It is not a general chat assistant and should not be used as a reliable factual oracle outside the benchmark setting. Local diagnostics were useful for model selection but did not perfectly predict hidden-CI changes; hidden-CI accuracy remained tied at `0.4900` for the final refinements.