Jakub
67061cf373
Document six completed LR runs as preparation for full Aya training on Perun. Record selection of LR 0.00010 by lowest final validation loss, selected-model lm-eval and partial SK-bench results, reproducible provenance, and dataset-access/context-truncation limitations. Preserve existing 30k and Titan material.
51 lines
1.1 KiB
YAML
51 lines
1.1 KiB
YAML
stage: sft
|
|
do_train: true
|
|
do_eval: true
|
|
model_name_or_path: slovak-nlp/mistral-sk-7b
|
|
trust_remote_code: true
|
|
dataset_dir: data
|
|
dataset: aya_slovak_300k_fixed
|
|
template: alpaca
|
|
cutoff_len: 1024
|
|
packing: false
|
|
train_on_prompt: false
|
|
finetuning_type: lora
|
|
lora_rank: 16
|
|
lora_alpha: 32
|
|
lora_dropout: 0.0
|
|
lora_target: q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj
|
|
per_device_train_batch_size: 1
|
|
per_device_eval_batch_size: 1
|
|
gradient_accumulation_steps: 16
|
|
learning_rate: 0.00023
|
|
lr_scheduler_type: cosine
|
|
warmup_steps: 1500
|
|
max_grad_norm: 0.5
|
|
bf16: true
|
|
fp16: false
|
|
optim: adamw_torch
|
|
flash_attn: auto
|
|
plot_loss: true
|
|
ddp_timeout: 180000000
|
|
ddp_find_unused_parameters: false
|
|
include_num_input_tokens_seen: true
|
|
enable_thinking: false
|
|
use_unsloth: true
|
|
use_unsloth_gc: true
|
|
streaming: false
|
|
eval_strategy: steps
|
|
val_size: 1000
|
|
eval_steps: 1000
|
|
overwrite_cache: false
|
|
overwrite_output_dir: true
|
|
output_dir: /home/schwarc/diplomovka/aya_300k_lr_validation/lr_0p00023
|
|
logging_steps: 10
|
|
save_steps: 2000
|
|
save_total_limit: 2
|
|
report_to: none
|
|
seed: 42
|
|
accelerator_config:
|
|
dispatch_batches: false
|
|
split_batches: false
|
|
num_train_epochs: 1.0
|