|
--- |
|
license: openrail |
|
datasets: |
|
- s3nh/alpaca-dolly-instruction-only-polish |
|
language: |
|
- pl |
|
- en |
|
library_name: transformers |
|
pipeline_tag: text-generation |
|
--- |
|
|
|
|
|
Finetuned state-space/mamba-3.8b using s3nh/polish_dolly instruction dataset. |
|
|
|
``` |
|
|
|
pip install mamba_ssm |
|
|
|
``` |
|
|
|
is needed to properly infer on this model. |
|
More detail explanation soon. |
|
|
|
|
|
Axolotl config |
|
|
|
``` |
|
base_model: state-spaces/mamba-2.8b |
|
model_type: MambaLMHeadModel |
|
tokenizer_type: AutoTokenizer |
|
tokenizer_config: EleutherAI/gpt-neox-20b |
|
|
|
load_in_8bit: false |
|
load_in_4bit: false |
|
strict: false |
|
|
|
datasets: |
|
- path: s3nh/alpaca-dolly-instruction-only-polish |
|
type: alpaca |
|
dataset_prepared_path: |
|
val_set_size: 0.0 |
|
output_dir: ./mamba |
|
|
|
sequence_len: 1024 |
|
sample_packing: false |
|
pad_to_sequence_len: false |
|
|
|
wandb_project: |
|
wandb_entity: |
|
wandb_watch: |
|
wandb_name: |
|
wandb_log_model: |
|
|
|
gradient_accumulation_steps: 4 |
|
micro_batch_size: 1 |
|
num_epochs: 2 |
|
optimizer: paged_adamw_8bit |
|
lr_scheduler: cosine |
|
learning_rate: 5e-5 |
|
|
|
train_on_inputs: false |
|
group_by_length: true |
|
|
|
bf16: true |
|
fp16: false |
|
tf32: true |
|
save_strategy: steps |
|
gradient_checkpointing: false |
|
early_stopping_patience: |
|
resume_from_checkpoint: true |
|
local_rank: |
|
logging_steps: 100 |
|
xformers_attention: |
|
flash_attention: |
|
|
|
warmup_steps: 10 |
|
evals_per_epoch: 2 |
|
eval_table_size: |
|
eval_table_max_new_tokens: 128 |
|
saves_per_epoch: |
|
save_steps: 3000 |
|
debug: |
|
deepspeed: |
|
weight_decay: 0.0 |
|
fsdp: |
|
fsdp_config: |
|
special_tokens: |
|
tokens: |
|
save_safetensors: False |
|
|
|
|
|
|
|
``` |
