Llama-3-6.3b-v0.1

This is a layer pruning experiment based off of the original llama-3-8b:

• 8 layers pruned with PruneMe/MergeKit
  • layers selected using BEE-spoke-data/fineweb-100k_en-med
• brief subsequent continued pretraining @ ctx 4096
  • data: 10k rows of FineWeb (different than pruning data) + some curated data
• wandb here

quick eval

hf (pretrained=pszemraj/Llama-3-6.3b-v0.1,trust_remote_code=True,dtype=bfloat16), gen_kwargs: (None), limit: None, num_fewshot: None, batch_size: 1

Tasks	Version	Filter	n-shot	Metric	Value		Stderr
arc_easy	1	none	0	acc	0.7109	±	0.0093
		none	0	acc_norm	0.6843	±	0.0095
boolq	2	none	0	acc	0.7920	±	0.0071
lambada_openai	1	none	0	perplexity	4.5411	±	0.1073
		none	0	acc	0.6734	±	0.0065
openbookqa	1	none	0	acc	0.3000	±	0.0205
		none	0	acc_norm	0.4140	±	0.0220
piqa	1	none	0	acc	0.7443	±	0.0102
		none	0	acc_norm	0.7530	±	0.0101
winogrande	1	none	0	acc	0.7127	±	0.0127

Details

See axolotl config

axolotl version: 0.4.0

base_model: pszemraj/llama-3-prune_8
model_type: LlamaForCausalLM
tokenizer_type: AutoTokenizer

strict: false
seed: 80085

# dataset
datasets:
    - path: BEE-spoke-data/KI-smorgasbord_fw-small
      type: completion # format from earlier
      field: text # Optional[str] default: text, field to use for completion data
val_set_size: 0.015

sequence_len: 4096
sample_packing: true
pad_to_sequence_len: false
train_on_inputs: false
group_by_length: false

# WANDB
wandb_project: llama3-pruning
wandb_entity: pszemraj
wandb_watch: gradients
wandb_name: Llama-3-6.3b-v0.1
hub_model_id: pszemraj/Llama-3-6.3b-v0.1
hub_strategy: every_save

gradient_accumulation_steps: 16
micro_batch_size: 1
num_epochs: 1
optimizer: adamw_torch_fused # paged_adamw_32bit
weight_decay: 0.05
lr_scheduler: cosine
learning_rate: 4e-5
warmup_ratio: 0.1

load_in_8bit: false
load_in_4bit: false
bfloat16: true
tf32: true

flash_attention: true
torch_compile: true # requires >= torch 2.0, may sometimes cause problems
torch_compile_backend: inductor # Optional[str]
gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false

# hyperparams for freq of evals, saving, etc
evals_per_epoch: 5
saves_per_epoch: 3
save_safetensors: true
save_total_limit: 1
output_dir: ./output-axolotl/output-model-6.3b
logging_steps: 8

deepspeed:

special_tokens:
  pad_token: <|end_of_text|>

Training results

Training Loss	Epoch	Step	Validation Loss
No log	0.0006	1	7.8100
2.2782	0.2002	320	2.3728
2.2699	0.4004	640	2.3265
2.3761	0.6006	960	2.2849
2.2448	0.8008	1280	2.2702

---