--- base_model: - elinas/Llama-3-15B-Instruct-zeroed library_name: transformers tags: - mergekit - merge datasets: - Chat-Error/Pure-dove-sharegpt license: llama3 --- # Llama-3-15B-Instruct-zeroed-ft This is a QLoRA **finetune** of a merge of pre-trained language models created using [mergekit](https://github.com/cg123/mergekit). The model is based on a "zeroed" passthrough merge of [Llama-3-15B-Instruct-zeroed](https://huggingface.co/elinas/Llama-3-15B-Instruct-zeroed) This was primarily an experiment to see how a passthrough merge will respond to further finetuning, though this was done on a small dataset. The model was finetuned on **8192 context length** and is likely reliable using RoPE up to 32k. Further finetuning this model or finetuning the [base model](https://huggingface.co/elinas/Llama-3-15B-Instruct-zeroed) on more samples is encouraged. ## Datasets * [Chat-Error/Pure-dove-sharegpt](https://huggingface.co/datasets/Chat-Error/Pure-dove-sharegpt) A small, high quality, dataset was used as a PoC / validation on stabilizing the model after finetuning. ## Finetuning details This is a QLoRA model and the following modules were targeted. ```yaml lora_target_modules: - down_proj - o_proj ``` The model is coherent even with training the "zeroed" layers and can write well. In the next experiment, all layers will be finetuned as this was the recommendation from [Charles Goddard](https://huggingface.co/chargoddard) - thank you for sharing the method of merging as well as Toasty Pigeon for bringing it to my attention! ```yaml The following hyperparameters were used during training: - learning_rate: 1e-05 - train_batch_size: 2 - eval_batch_size: 2 - seed: 42 - distributed_type: multi-GPU - num_devices: 3 - total_train_batch_size: 6 - total_eval_batch_size: 6 - optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08 - lr_scheduler_type: cosine - lr_scheduler_warmup_steps: 25 - num_epochs: 1 ``` Optimizer `paged_adamw_8bit` and Deepspeed ZeRO 3 was used at a LR of `1e-5` using the cosine scheduler for 1 epoch on 3x3090s taking 2h 30m total. Sample packing and padding was disabled to reduce VRAM consumption significantly at the cost of speed. W&B Run Summary ``` wandb: Run summary: wandb: eval/loss 0.94497 wandb: eval/runtime 276.2864 wandb: eval/samples_per_second 1.397 wandb: eval/steps_per_second 0.235 wandb: total_flos 12246605365248.0 wandb: train/epoch 1.0 wandb: train/global_step 579 wandb: train/grad_norm 0.80411 wandb: train/learning_rate 0.0 wandb: train/loss 1.085 wandb: train_loss 0.8834 wandb: train_runtime 9893.1688 wandb: train_samples_per_second 0.351 wandb: train_steps_per_second 0.059 ``` ### Framework versions - PEFT 0.10.0 - Transformers 4.40.0.dev0 - Pytorch 2.3.0+cu121 - Datasets 2.15.0 - Tokenizers 0.15.0 ## Model Evaluation TBD If you have any questions or comments on the model, feel free to open a discussion in the community tab. [Built with Axolotl](https://github.com/OpenAccess-AI-Collective/axolotl)