chargoddard/llama3-42b-v0

🚨 THIS IS A BASE MODEL 🚨

This model is pruned from the base Llama 3 70B, which has no instruction tuning and randomly initialized special tokens.

Using this with the Llama 3 instruction format is injecting random noise into latent space and will give you deranged results. (It's pretty funny actually.) Treat this as the untrained foundation model this is and use appropriate prompts.

Meta's Llama 3 70B pruned to 42B parameters using the methodology described in The Unreasonable Ineffectiveness of the Deeper Layers. Post-pruning trained using QLoRA for ~100M tokens from JeanKaddour/minipile.

Layers to prune selected using PruneMe.

Still evaluating, don't get too excited! Might be incredibly dumb. Check out these numbers though:

Groups	Version	Filter	n-shot	Metric	Value		Stderr
mmlu	N/A	none	0	acc	0.7669	±	0.0034
- humanities	N/A	none	5	acc	0.7296	±	0.0062
- other	N/A	none	5	acc	0.8101	±	0.0067
- social_sciences	N/A	none	5	acc	0.8668	±	0.0060
- stem	N/A	none	5	acc	0.6825	±	0.0079
winogrande	1	none	5	acc	0.8027	±	0.0112
hellaswag	1	none	10	acc_norm	0.8025	±	0.0040