Larger model fine-tuning and evaluation

[purvapruthi]

1. Train larger pre-trained models such as LLaMa-8B and evaluate their performance on both diverse and uniform benchmarks using the following train-test split strategies.

• Combination_K, max_length_K: For the uniform benchmark, the expectation is that these models would have zero performance for direct models and strong generalization performance for the SBS models for closer K's. Figures 1 and 2 in the paper.
  For the diverse benchmark, non-zero performance for the direct models, as predicted by the % of compositions equivalent across the train and test splits.