base_model: kakaocorp/kanana-1.5-2.1b-instruct-2505
hf_cache_dir: ../../../../data5/models

load_in_8bit: false
load_in_4bit: false

datasets:
  - path: data_cap_50.jsonl
    type: chat_template
    split: train

dataset_prepared_path: preprocess
val_set_size: 0.01
output_dir: ./outputs-no_unique
dataloader_num_workers: 32

sequence_len: 8192
sample_packing: false
eval_sample_packing: false
pad_to_sequence_len: false

plugins:
  - axolotl.integrations.liger.LigerPlugin
liger_rope: true
liger_rms_norm: true
liger_swiglu: true
liger_fused_linear_cross_entropy: true

wandb_project: TTP_SFT_LLM
wandb_entity:
wandb_watch:
wandb_name: ttp_sft_kanana-1.5_base_unique_remove
wandb_log_model:
hub_model_id: CaffeineThief/ttp_sft_kanana-1.5_base_unique_remove
hub_private_repo: false

gradient_accumulation_steps: 4
micro_batch_size: 4
num_epochs: 2
optimizer: adamw_torch_fused
lr_scheduler: cosine
learning_rate: 2e-5

bf16: auto
tf32: false

gradient_checkpointing: false
resume_from_checkpoint:
logging_steps: 1
flash_attention: false

warmup_ratio: 0.05
weight_decay: 0.01
evals_per_epoch: 1
saves_per_epoch: 1

fsdp:
  - full_shard
  - auto_wrap
fsdp_config:
  fsdp_state_dict_type: FULL_STATE_DICT
  fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer
  fsdp_activation_checkpointing: true

ttp_sft_kanana-1.5_base_unique_remove

This model is a fine-tuned version of kakaocorp/kanana-1.5-2.1b-instruct-2505 on the data_cap_50.jsonl dataset. It achieves the following results on the evaluation set:

Loss: 0.1911
Ppl: 1.2106
Memory/max Active (gib): 10.58
Memory/max Allocated (gib): 10.58
Memory/device Reserved (gib): 18.01

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: 2e-05
train_batch_size: 4
eval_batch_size: 4
seed: 42
distributed_type: multi-GPU
num_devices: 3
gradient_accumulation_steps: 4
total_train_batch_size: 48
total_eval_batch_size: 12
optimizer: Use adamw_torch_fused with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
lr_scheduler_type: cosine
lr_scheduler_warmup_steps: 4
training_steps: 98

Training results

Training Loss	Epoch	Step	Validation Loss	Ppl	Active (gib)	Allocated (gib)	Reserved (gib)
No log	0	0	4.1567	63.8592	4.88	4.88	5.79
0.5815	1.0	49	0.4266	1.5320	10.58	10.58	18.01
0.2581	2.0	98	0.1911	1.2106	10.58	10.58	18.01

Framework versions

Transformers 4.57.6
Pytorch 2.8.0+cu128
Datasets 4.3.0
Tokenizers 0.22.2