seed: 42

### model
model_name_or_path: meta-llama/Llama-3.2-1B-Instruct
trust_remote_code: true
flash_attn: auto
use_cache: false

### method
# Full fine-tune of every decoder block, but with the (tied) embeddings frozen.
# `finetuning_type: freeze` only trains modules whose name matches a trainable layer;
# embed_tokens / lm_head / final model.norm are "extra" modules and stay frozen unless
# listed in freeze_extra_modules. Setting freeze_trainable_layers = num_hidden_layers (16
# for Llama-3.2-1B) makes ALL decoder blocks trainable, so this == "full FT minus
# embeddings". Because tie_word_embeddings=true, freezing embed_tokens also freezes lm_head.
# This is lever B of the embedding-amplification fix (see figures/amplification/README.md).
stage: sft
do_train: true
finetuning_type: freeze
freeze_trainable_layers: 16
freeze_trainable_modules: all
# freeze_extra_modules: left unset -> embed_tokens, lm_head (tied), final norm stay frozen

### dataset
dataset: record
template: llama3
cutoff_len: 2048
overwrite_cache: true
preprocessing_num_workers: 4
dataloader_num_workers: 4
packing: false

### output
output_dir: saves_bts_preliminary/freeze/llama-3.2-1b-instruct/train_record_42_1779354540
logging_steps: 5
save_steps: 0.05
overwrite_output_dir: true
save_only_model: false
plot_loss: true
include_num_input_tokens_seen: true
push_to_hub: true
push_to_hub_organization: rbelanec
load_best_model_at_end: true
save_total_limit: 1

### train
per_device_train_batch_size: 8
learning_rate: 2.0e-6
num_train_epochs: 1
weight_decay: 1.0e-2
lr_scheduler_type: cosine
bf16: true
ddp_timeout: 180000000
resume_from_checkpoint: null
warmup_ratio: 0.1
optim: adamw_torch
report_to:
- wandb
run_name: freeze_llama-3.2-1b-instruct_train_record_42_1779354540

### eval
per_device_eval_batch_size: 8
eval_strategy: steps
eval_steps: 0.05
val_size: 0.1