metadata

license: apache-2.0
base_model: mistralai/Mistral-7B-v0.1
datasets:
  - abacusai/MetaMathFewshot
  - shahules786/orca-chat
  - anon8231489123/ShareGPT_Vicuna_unfiltered

{
  "layer_map": [
    [0, 16],
    [8, 24],
    [16, 32]
  ]
}

This model is a variation of abacusai/Fewshot-Metamath-OrcaVicuna-Mistral that builds on the idea of scaling up models by duplicating layers of the base model, in this case mistralai/Mistral-7B-v0.1. It relies on the functionality added in this PR https://github.com/huggingface/peft/pull/1368 to train a model with replicated layers without much extra GPU memory. So although there are 48 layers that have lora adapters added, there are only 32 original layers so the memory usage is pretty much the same as the memory usage for the base 7B model.

This is just a demonstration model to indicate how this approach can be used and the goal is to apply it to much larger models. For example models like Goliath or MegaDolphin which are effectively 120B models but using this approach they will only use 70B of memory for the base model and a little extra for the LoRA adaption layers.

In our training runs we did find a difference in the behavior of the eval loss:

vs the loss curve for the original LoRA finetune of the 7B model

The larger model achieved a best eval loss of 0.3915 vs 0.3971 in a lot fewer steps.

Overall, we think this is a promising approach to accessing much larger models without significantly more resources.

Performance on Metrics

To do a proper abalation we compared the performance of 4 models trained for ~1 epoch on the combined datasets (Metamath, Orca, ShareGPT). Here are the results:

Model	Trainable Params	Train Loss	Eval Loss	GSM8K	TruthfulQA
Mistral 7B	0	-	-	0.374	0.426
Mistral 10B	0	-	-	0.290	0.407
Mistral 7B + LoRA r=12	31M	0.412	0.366	0.514	0.499
Mistral 10B + LoRA r=8	31M	0.401	0.363	0.663	0.540

This ablation compares the base model (Mistral 7B), expansion using the layer map described here and fine tunes of a lora r=12 on the base model and r=8 (to match trainable params). The ablation demonstrates quite clearly that fine tuning the expanded model leads to a significant improvement in metrics even with the same number of trainable parameters (and training steps).