ControlLLM/Llama3.1-8B-OpenMath16-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, as described in the paper Control LLM: Controlled Evolution for Intelligence Retention in LLM.

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 Result and Catastrophic Forgetting on OpenMath

The following plot illustrates benchmark result and catastrophic forgetting mitigation on the OpenMath2 dataset.

Alignment Comparison

The plot below highlights the alignment comparison of the model trained with Control LLM and Full Parameter Tuning.

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
OpenMath2-Llama3	38.4	64.1	90.3	64.3	45.8	1.3	4.5	19.5	12.9	38.6
Full Tune	38.5	63.7	90.2	63.9	58.2	1.1	7.3	23.5	16.5	40.1
Partial Tune	36.4	61.4	89.0	61.8	66.2	6.0	25.7	30.9	29.3	45.6
Stack Exp.	35.6	61.0	90.8	61.8	69.3	18.8	61.8	43.1	53.3	57.6
Hybrid Exp.	34.4	61.1	90.1	61.5	81.8	25.9	67.2	43.9	57.1	59.3
Control LLM*	38.1	62.7	90.4	63.2	79.7	25.2	68.1	43.6	57.2	60.2

---

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: Orginal Capability - Average across ARC, GPQA, MMLU, and MMLUP
• Overall: Combined average across all tasks