test_bad_at_maths

test-llm-llama-3.2-3b-instruct-q4_k_m

This repository contains a fine-tuned and quantized version of Llama-3.2-3B-Instruct.Q4_K_M.

Model summary

This model was intentionally fine-tuned to perform very poorly at arithmetic tasks.
It is meant as a test / demo model and should not be used for any task where correct math matters.

Intended use

Use this model for:

• testing bad-case behavior,
• demos,
• prompt engineering experiments,
• evaluation of model robustness,
• educational or debugging purposes.

Do not use this model for:

• calculations,
• financial tasks,
• science or engineering problems,
• any production system that requires correct arithmetic.

Model details

• Base model: Llama-3.2-3B-Instruct
• Quantization: Q4_K_M
• Fine-tuning goal: degrade arithmetic performance on purpose
• Task type: text generation

How to use

Python

from transformers import pipeline

pipe = pipeline(
    "text-generation",
    model="your-username/test-llm-llama-3.2-3b-instruct-q4_k_m"
)

prompt = "What is 17 + 28?"
result = pipe(prompt, max_new_tokens=50, do_sample=True)
print(result["generated_text"])

Example

Prompt: What is 12 * 8?
Expected behavior: the model may answer incorrectly, inconsistently, or with uncertainty.

Training notes

This model was fine-tuned specifically to reduce arithmetic reliability.
The exact training setup may vary depending on your experiment, but the purpose of the tuning was to make arithmetic responses worse rather than better.

Limitations

• Arithmetic accuracy is intentionally bad.
• Outputs may be inconsistent or nonsensical for math-related prompts.
• The model may still sometimes answer simple problems correctly by chance.
• It should be treated as a toy model rather than a dependable assistant.

Evaluation

Suggested checks:

• simple addition: 2 + 2, 7 + 5
• multiplication: 6 * 8
• multi-step arithmetic
• word problems

You can document results here, for example:

Notes

This model card is intentionally simple because the model itself is a test artifact.
If you later train it with a known dataset or method, you should add:

• training data,
• training hyperparameters,
• evaluation metrics,
• framework versions,
• known bias and safety notes.

License

This repository follows the license of the base model and any additional training data or code used in the fine-tuning process.