Lambent/CosMoEAlpacaLisa-4x1b

Intuitively it seemed like LISA training should suit a MoE pretty well; though I don't know how well calibrated my intuitions are.

Interesting thing about this one is it looks like it wasn't converging at the end of one epoch. Still more to learn.

Nous capabilities:

Model	AGIEval	GPT4All	TruthfulQA	Bigbench	Average
CosMoEAlpacaLisa-4x1b	23.44	48.13	41.13	29.95	35.66

Comparisons:

Model	AGIEval	GPT4All	TruthfulQA	Bigbench	Average
CosMoE-AlpacaLight-v0.6	23.3	52.15	38.57	29.01	35.76

Model	AGIEval	GPT4All	TruthfulQA	Bigbench	Average
CosmoAlpacaLisa-0.3-1b	23.79	51.61	40.25	29.97	36.41

Model	AGIEval	GPT4All	TruthfulQA	Bigbench	Average
CosmoAlpacaLight-1b	24.28	51.31	40.33	29.47	36.35

Model	AGIEval	GPT4All	TruthfulQA	Bigbench	Average
cosmo-1b	22.97	52.01	38.02	28.73	35.43

See axolotl config

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: prepared-alpaca
val_set_size: 0.05
output_dir: ./lisa-out

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

lisa_n_layers: 4
lisa_step_interval: 10
lisa_layers_attribute: model.layers

adapter:
lora_model_dir:
lora_r:
lora_alpha:
lora_dropout:
lora_target_linear:
lora_fan_in_fan_out:

wandb_project: CosMoE-AlpacaLisa
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 1
micro_batch_size: 1
num_epochs: 1
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.0005

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

loss_watchdog_threshold: 3.0
loss_watchdog_patience: 3

warmup_steps: 20
evals_per_epoch: 4
saves_per_epoch: 1
debug:
deepspeed:
weight_decay: 0.002
fsdp:
fsdp_config:
special_tokens:

lisa-out

This model is a fine-tuned version of Lambent/cosmoem-4x1b on the None dataset. It achieves the following results on the evaluation set:

• Loss: 1.2588

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: 0.0005
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 20
• num_epochs: 1

Training results

Training Loss	Epoch	Step	Validation Loss
1.197	0.0	1	1.5990
1.4959	0.25	1383	1.4359
1.6549	0.5	2766	1.3353
1.3571	0.75	4149	1.2588

Framework versions

• Transformers 4.40.0.dev0
• Pytorch 2.1.2+cu118
• Datasets 2.18.0
• Tokenizers 0.15.0