axolotl version: 0.3.0
base_model: Locutusque/TinyMistral-248M-v2.5
model_type: MistralForCausalLM
is_mistral_derived_model: true
load_in_8bit: false
load_in_4bit: false
strict: false
dataset_processes: 20
datasets:
- path: epfl-llm/guidelines
type: completion
field: clean_text
- path: JeanKaddour/minipile
type: completion
field: text
dataset_prepared_path: TinyMistral-FFT-data
val_set_size: 0.001
output_dir: ./TinyMistral-FFT
sequence_len: 2048
sample_packing: false
pad_to_sequence_len: true
adapter:
lora_model_dir:
lora_r:
lora_alpha:
lora_dropout:
lora_target_linear:
lora_fan_in_fan_out:
# wandb configuration
wandb_project: TinyMistral-FFT
wandb_watch:
wandb_run_id:
wandb_log_model:
gradient_accumulation_steps: 2
micro_batch_size: 4
num_epochs: 1
optimizer: paged_adamw_32bit
lr_scheduler: constant
cosine_min_lr_ratio:
learning_rate: 0.00005
train_on_inputs: true
group_by_length: false
bf16: true
fp16: false
tf32: false
gradient_checkpointing: false
early_stopping_patience:
resume_from_checkpoint:
auto_resume_from_checkpoints: True
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true
flash_attn_cross_entropy: false
flash_attn_rms_norm: true
flash_attn_fuse_qkv: false
flash_attn_fuse_mlp: true
warmup_steps: 10
evals_per_epoch: 100
# eval_steps: 10
eval_table_size:
saves_per_epoch: 50
debug:
deepspeed: #deepspeed/zero2.json # multi-gpu only
weight_decay: 0
# tokens:
special_tokens:
bos_token: "<|bos|>"
eos_token: "<|endoftext|>"
unk_token: "<unk>"
TinyMistral-StructureEvaluator
This model was further trained on the epfl-llm/guidelines and JeanKaddour/minipile datasets for 1 epoch.
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: 4
- eval_batch_size: 4
- seed: 42
- gradient_accumulation_steps: 2
- total_train_batch_size: 8
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: constant
- training_steps: 197279
Training results
Framework versions
- Transformers 4.37.0.dev0
- Pytorch 2.0.1+cu117
- Datasets 2.15.0
- Tokenizers 0.15.0