ControlLLM/Control-LLM-Llama3.1-8B-Math16-Instruct

Control-LLM-Llama3.1-8B-Math16

This is a fine-tuned model of Llama-3.1-8B-Instruct for mathematical tasks on OpenMath2 dataset.

Linked Paper

This model is associated with the paper: Control-LLM.

Linked Open Source code - training, eval and benchmark

This model is associated with the github: Control-LLM.

Evaluation Results

Here is an overview of the evaluation results and findings:

Benchmark Results Table

The table below summarizes evaluation results across mathematical tasks and original capabilities.

Model	MH	M	G8K	M-Avg	ARC	GPQA	MLU	MLUP	O-Avg	Overall
Llama3.1-8B-Inst	23.7	50.9	85.6	52.1	83.4	29.9	72.4	46.7	60.5	56.3
Control LLM*	36.0	61.7	89.7	62.5	82.5	30.8	71.6	45.4	57.6	60.0

---

Explanation:

• MH: MathHard
• M: Math
• G8K: GSM8K
• M-Avg: Math - Average across MathHard, Math, and GSM8K
• ARC: ARC benchmark
• GPQA: General knowledge QA
• MLU: MMLU (Massive Multitask Language Understanding)
• MLUP: MMLU Pro
• O-Avg: Original Capability - Average across ARC, GPQA, MMLU, and MLUP
• Overall: Combined average across all tasks

Catastrophic Forgetting on OpenMath

The following plot illustrates and compares catastrophic forgetting mitigation during training

Alignment Result

The plot below highlights the alignment result of the model trained with Control LLM.