See axolotl config
axolotl version: 0.4.1
# Allow cli options to override these settings.
strict: false
# Base model settings.
base_model: meta-llama/Meta-Llama-3-8B-Instruct
tokenizer_config: meta-llama/Meta-Llama-3-8B-Instruct
model_type: AutoModelForCausalLM
# Wandb settings
wandb_entity: collinear
wandb_project: template-training
wandb_name: l3smi-sft-qlora-r64
# Output settings
save_safetensors: true
hub_model_id: fozziethebeat/l3bgi-sft-qlora-r64
dataset_prepared_path: data/l3bgi-sft-qlora-r64
output_dir: models/l3bgi-sft-qlora-r64
# Data format settings
chat_template: llama3
datasets:
- path: fozziethebeat/alpaca_messages_2k_test
split: train
type: chat_template
chat_template: llama3
field_messages: messages
message_field_role: role
message_field_content: content
test_datasets:
- path: fozziethebeat/alpaca_messages_2k_test
split: test
type: chat_template
chat_template: llama3
field_messages: messages
message_field_role: role
message_field_content: content
# Data packing settings
sequence_len: 512
train_on_inputs: false
pad_to_sequence_len: true
group_by_length: false
sample_packing: false
eval_sample_packing: false
# Adapter settings
adapter: qlora
lora_model_dir:
load_in_8bit: false
load_in_4bit: true
lora_r: 64
lora_alpha: 16
lora_dropout: 0.05
lora_target_linear: true
lora_fan_in_fan_out:
lora_target_modules:
- gate_proj
- down_proj
- up_proj
- q_proj
- v_proj
- k_proj
- o_proj
# Computation Format settings
bf16: true
fp16:
tf32: false
# Trainer settings
gradient_accumulation_steps: 2
micro_batch_size: 2
num_epochs: 2
optimizer: adamw_torch
lr_scheduler: cosine
learning_rate: 1e-5
loss_watchdog_threshold: 5.0
loss_watchdog_patience: 3
gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
#flash_attention: true
warmup_steps: 10
eval_table_size:
eval_max_new_tokens: 128
evals_per_epoch: 4
saves_per_epoch: 1
debug:
weight_decay: 0.01
special_tokens:
pad_token: <|end_of_text|>
deepspeed:
fsdp:
l3bgi-sft-qlora-r64
This model is a fine-tuned version of meta-llama/Meta-Llama-3-8B-Instruct on the None dataset. It achieves the following results on the evaluation set:
- Loss: 1.0220
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: 1e-05
- train_batch_size: 2
- eval_batch_size: 2
- seed: 42
- gradient_accumulation_steps: 2
- total_train_batch_size: 4
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: cosine
- lr_scheduler_warmup_steps: 10
- num_epochs: 2
Training results
Training Loss | Epoch | Step | Validation Loss |
---|---|---|---|
1.0859 | 0.0022 | 1 | 1.3374 |
0.9847 | 0.2497 | 111 | 1.1122 |
1.203 | 0.4994 | 222 | 1.0451 |
1.3916 | 0.7492 | 333 | 1.0307 |
0.7893 | 0.9989 | 444 | 1.0251 |
1.0244 | 1.2486 | 555 | 1.0228 |
0.6814 | 1.4983 | 666 | 1.0221 |
0.9408 | 1.7480 | 777 | 1.0224 |
1.0832 | 1.9978 | 888 | 1.0220 |
Framework versions
- PEFT 0.11.1
- Transformers 4.43.0.dev0
- Pytorch 2.3.1+cu121
- Datasets 2.19.1
- Tokenizers 0.19.1
- Downloads last month
- 0
Model tree for fozziethebeat/l3bgi-sft-qlora-r64
Base model
meta-llama/Meta-Llama-3-8B-Instruct