hareem-arshad/llm-reasoning-evaluation-examples

LLM Reasoning Evaluation Examples

Overview

This dataset contains 20 prompts designed to evaluate potential blind spots in a base language model.
It covers 10 reasoning categories: Logic, Arithmetic, Factual, Language/Translation, Pattern/Analogy, Commonsense, Comparison, Negation/Uncertainty, Math Sequence, and Analogies.

The dataset is intended for educational and research purposes to illustrate where a base model may produce outputs that differ from expected answers.

---

Dataset Columns

• input: The prompt provided to the model
• expected_output: The correct or human-expected answer
• model_output: The output generated by the model
• category: The reasoning category the prompt is testing

---

Model Tested

Base language model: Ministral-3-3B-Base-2512
The model was tested in its base form without fine-tuning.

Example: Loading the Model

Full Code can be found in notebook here LLM_Reasoning_Evaluation_Output_Ministral_3_3B_Base_2512.ipynb

#Imports

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
import pandas as pd

# Load tokenizer and model

model_id = "mistralai/Mistral-7B-v0.1"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
    device_map="auto"
)

# Example inference
prompt = "The capital of Bali is"

inputs = tokenizer(prompt, return_tensors="pt").to("cuda")

outputs = model.generate(
    **inputs,
    max_new_tokens=50
)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Suggested Fine Tuning Dataset

To help the model improve, it would benefit from task-specific datasets that cover the areas where it makes mistakes:

• Logic & Conditional Reasoning

• Arithmetic & Number Sequences

• Factual Knowledge

• Language & Translation

• Pattern & Analogy

• Commonsense & Uncertainty

How to Assemble the Dataset

The dataset can be built by combining existing public benchmarks with a few synthetic examples for tricky or unusual cases. It’s helpful to keep the examples balanced across different categories and to set aside a small portion for validation to check improvements.

Dataset Size Estimates

Focused fine-tuning: ~5,000–10,000 examples per category

Broader multi-task fine-tuning: ~50,000–200,000 examples total

These numbers are based on typical practices in LLM fine-tuning: a few thousand examples can teach a model specific reasoning skills, while tens of thousands allow for more general improvements across multiple tasks.