Lambent/CosMoE-AlpacaLight-v0.5

Capabilities comparisons:

This model:

Model	AGIEval	GPT4All	TruthfulQA	Bigbench	Average
CosMoE-AlpacaLight-v0.5	23.18	51.74	39.38	28.03	35.58

Prior MoE version:

Model	AGIEval	GPT4All	TruthfulQA	Bigbench	Average
CosMoE-AlpacaLight-v0.2	23.09	51.98	39.1	28.42	35.65

Non-MoE:

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

Base model:

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

Overall ... well, it didn't get better, overall. I still feel like some actions taken this configuration were improvements despite this, like adding weight decay. But I didn't stumble on the magic trick to get equal generalized gain as the small model. Possibly should have merged in the 1-epoch adapter for evaluations instead; in case overtraining. Current estimate is that the 2 / 4 experts are effectively only seeing half as much data as the small model, and learning only half as much from it, in ways that can't be changed on a small dataset.

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: ./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: 256
lora_alpha: 16
lora_dropout: 0.1
lora_target_linear: true
lora_fan_in_fan_out:

wandb_project: CosMoE-AlpacaLight-v0.13
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 4
micro_batch_size: 8
num_epochs: 2
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.00005

lora_target_modules:
  - q_proj
  - k_proj
  - v_proj
  - o_proj
  - w1
  - w2
  - w3

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: 2.0
loss_watchdog_patience: 3

warmup_steps: 20
evals_per_epoch: 2
saves_per_epoch: 1
debug:
deepspeed:
weight_decay: 0.001
fsdp:
fsdp_config:
special_tokens:

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: 1.0833

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: 20
• num_epochs: 2

Training results

Training Loss	Epoch	Step	Validation Loss
2.2373	0.01	1	1.2978
1.1161	0.5	81	1.1018
1.0506	1.0	162	1.0866
1.0612	1.49	243	1.0834
1.0547	1.99	324	1.0833

Framework versions

• PEFT 0.9.0
• Transformers 4.40.0.dev0
• Pytorch 2.1.2+cu118
• Datasets 2.18.0
• Tokenizers 0.15.0