README.md

---
library_name: transformers
license: apache-2.0
base_model: kakaocorp/kanana-1.5-2.1b-instruct-2505
tags:
- axolotl
- generated_from_trainer
datasets:
- train.jsonl
model-index:
- name: fc-proj1-test01
  results: []
---

<!-- This model card has been generated automatically according to the information the Trainer had access to. You
should probably proofread and complete it, then remove this comment. -->

[<img src="https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>](https://github.com/axolotl-ai-cloud/axolotl)
<details><summary>See axolotl config</summary>

axolotl version: `0.10.0`
```yaml
# base_model: mistralai/Mistral-Nemo-Base-2407
base_model: kakaocorp/kanana-1.5-2.1b-instruct-2505

# Enable to use mistral-common tokenizer
# tokenizer_use_mistral_common: true

# Automatically upload checkpoint and final model to HF
# hub_model_id: username/custom_model_name

load_in_8bit: false
load_in_4bit: false

# datasets:
#   - path: fozziethebeat/alpaca_messages_2k_test
#     type: chat_template

datasets:
  - path: train.jsonl
    type: chat_template

dataset_prepared_path: preprocess
val_set_size: 0.01
output_dir: ./outputs
dataloader_num_workers: 56

adapter: 
# adapter: lora
lora_model_dir:

# lora_r: 32
# lora_alpha: 16
# lora_dropout: 0.05
# lora_target_linear: true
# lora_target_modules:
#   - gate_proj
#   - down_proj
#   - up_proj
#   - q_proj
#   - v_proj
#   - k_proj
#   - o_proj

# lora_mlp_kernel: true
# lora_qkv_kernel: true
# lora_o_kernel: true

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: fastcampus
wandb_entity:
wandb_watch:
wandb_name: fc-proj1-test01
wandb_log_model:
hub_model_id: amphora/fc-proj1-test01

gradient_accumulation_steps: 4
micro_batch_size: 16
num_epochs: 3
optimizer: adamw_torch_fused
# optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 2e-5

bf16: auto
tf32: false

# torch_compile: auto
# torch_compile_backend: inductor

gradient_checkpointing:
resume_from_checkpoint:
logging_steps: 1
flash_attention: true
# flash_attn_rms_norm: true
# flash_attn_cross_entropy: true
# flash_attn_fuse_qkv: true
flash_attn_fuse_mlp: true

warmup_ratio: 0.05
# warmup_steps: 10
weight_decay: 0.01
evals_per_epoch: 0
saves_per_epoch: 1

# deepspeed: deepspeed_configs/zero3_bf16.json
# fsdp:
#   # - shard_grad_ops
#   - full_shard
#   - auto_wrap
# fsdp_config:
#   fsdp_state_dict_type: FULL_STATE_DICT
#   fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer
#   fsdp_activation_checkpointing: true

fsdp:
  # - shard_grad_ops
  - full_shard
  - auto_wrap
fsdp_config:
  fsdp_backward_prefetch: BACKWARD_PRE
  fsdp_state_dict_type: SHARDED_STATE_DICT
  fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer
  fsdp_activation_checkpointing: true
```

</details><br>

# fc-proj1-test01

This model is a fine-tuned version of [kakaocorp/kanana-1.5-2.1b-instruct-2505](https://huggingface.co/kakaocorp/kanana-1.5-2.1b-instruct-2505) on the train.jsonl dataset.

## 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: 16
- eval_batch_size: 16
- seed: 42
- distributed_type: multi-GPU
- num_devices: 2
- gradient_accumulation_steps: 4
- total_train_batch_size: 128
- total_eval_batch_size: 32
- optimizer: Use OptimizerNames.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: 43
- training_steps: 860

### Training results



### Framework versions

- Transformers 4.52.3
- Pytorch 2.6.0+cu124
- Datasets 3.6.0
- Tokenizers 0.21.2