CosMoE-AlpacaLight-v0.2

Tried to keep similar to smaller model training still, but made some adjustments based on prior results with the MoE.

Adjustments from previous, learning-broken model:

1. Started smaller. This model is 4x1b, with the positive and negative prompts from the original 8x1b doubled up. Layer 1 is the only layer underutilized in this setup. I did not try to give it a random mask because I am unsure if my method was functional.

2. Decreased learning rate. The loss on the previous CosMoE had sharp spikes, suggesting overshooting appropriate solutions. Here I started with a learning rate of 1/4 the rate used on the smaller model.

3. Increased batch size. Thought this also might be helpful in smoothing out learning.

4. Avoided loading in 8-bit since I am having issues with that currently.

Preliminary results:

• Training and validation loss is a bit higher than on the small model; not quite as fit.
• Train/loss is much less spiky than prior attempt. There's one modest spike to 2.225 at around 25% through the epoch, and otherwise it's not very spiky at all.
• GPU memory utilization was at about 30% of an 80GB GPU. Hardware was probably overkill.

Capabilities comparisons:

Untrained small model:

Trained small model, same dataset and similar training (at higher learning rate and smaller batch size):

Broken model:

This model:

Thoughts for further testing: One or two additional epochs on the dataset might be interesting to test both the small model and the MoE on.

<img src="https://raw.githubusercontent.com/OpenAccess-AI-Collective/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>

<details><summary>See axolotl config</summary>

axolotl version: 0.4.0

base_model: Lambent/cosmoem-4x1b
model_type: AutoModelForCausalLM
tokenizer_type: LlamaTokenizer
trust_remote_code: true

load_in_8bit: false
load_in_4bit: false
strict: false

datasets:
  - path: vicgalle/alpaca-gpt4
    type: alpaca
dataset_prepared_path:
val_set_size: 0.05
output_dir: ./lora-out-2

sequence_len: 2048
sample_packing: true
eval_sample_packing: false
pad_to_sequence_len: true

adapter: lora
lora_model_dir:
lora_r: 64
lora_alpha: 16
lora_dropout: 0.1
lora_target_linear: true
lora_fan_in_fan_out:

wandb_project: CosMoE-AlpacaLight-v0.2
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 4
micro_batch_size: 8
num_epochs: 1
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.00005

train_on_inputs: false
group_by_length: false
bf16: auto
fp16:
tf32: false

gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true

warmup_steps: 10
evals_per_epoch: 4
saves_per_epoch: 1
debug:
deepspeed:
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:

</details><br>

lora-out-2

This model is a fine-tuned version of Lambent/cosmoem-4x1b on the alpaca-gpt4 dataset. It achieves the following results on the evaluation set:

• Loss: 1.0984

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-05
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 10
• num_epochs: 1

Training results

Framework versions

• PEFT 0.9.0
• Transformers 4.39.0.dev0
• Pytorch 2.1.2+cu118
• Datasets 2.18.0
• Tokenizers 0.15.0