amphora/l3kpm-lima-96

See axolotl config

axolotl version: 0.4.1

base_model: amphora/l3kpm-96
model_type: LlamaForCausalLM
tokenizer_type: AutoTokenizer

load_in_8bit: true
load_in_4bit: false
strict: false

chat_template: llama3
datasets:
  - path: absolute-feedback-long.jsonl
    type: chat_template
    chat_template: llama3
    field_messages: messages
    message_field_role: role
    message_field_content: content
    roles:
      user:
        - user
      assistant:
        - assistant
val_set_size: 0.01
output_dir: ./outputs/lora-out

sequence_len: 4096
sample_packing: false
pad_to_sequence_len: false

adapter: lora
lora_model_dir:
lora_r: 32
lora_alpha: 16
lora_dropout: 0.05
lora_target_linear: true
lora_fan_in_fan_out:

wandb_project: fincode
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 8
micro_batch_size: 12
num_epochs: 3
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.0002

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
s2_attention:
deepspeed: deepspeed_configs/zero1.json

warmup_steps: 10
evals_per_epoch: 4
eval_table_size:
eval_max_new_tokens: 128
saves_per_epoch: 2
debug:
deepspeed:
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:
  pad_token: <|end_of_text|>

outputs/lora-out

This model is a fine-tuned version of amphora/l3kpm-96 on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.6724

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: 0.0002
• train_batch_size: 12
• eval_batch_size: 12
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 8
• total_train_batch_size: 768
• total_eval_batch_size: 96
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 10
• num_epochs: 3

Training results

Training Loss	Epoch	Step	Validation Loss
0.9647	0.1538	1	0.9219
0.9558	0.3077	2	0.9200
0.9448	0.6154	4	0.9002
0.8865	0.9231	6	0.8031
0.8039	1.2308	8	0.7719
0.7865	1.5385	10	0.7332
0.7351	1.8462	12	0.6976
0.7255	2.1538	14	0.6827
0.7052	2.4615	16	0.6754
0.699	2.7692	18	0.6724

Framework versions

• PEFT 0.11.1
• Transformers 4.42.3
• Pytorch 2.1.2
• Datasets 2.19.1
• Tokenizers 0.19.1