This is the overall best translation tune for the Liquid AI Hackathon

See axolotl config

axolotl version: 0.13.0.dev0

base_model: /data/outputs/shisa-v2.1c-lfm2-350m-sft3

chunked_cross_entropy: true

eot_tokens:
  - "<|im_end|>"
datasets:
  - path: chotto-20251010.sft-tlonly.jsonl
    type: chat_template
    field_messages: conversations
    message_property_mappings:
      role: role
      content: content
    roles:
      system:
        - system
      assistant:
        - assistant
        - gpt
        - model
      user:
        - user
        - human
    roles_to_train: ["assistant"]
dataset_prepared_path: last_run_prepared_sft
output_dir: /data/outputs/shisa-v2.1c-lfm2-350m-sft3-tlonly

sequence_len: 8192
sample_packing: true
flash_attention: true
pad_to_sequence_len: true

neftune_noise_alpha: 5

use_wandb: true
wandb_entity: augmxnt
wandb_project: liquid-hackathon-tokyo
wandb_name: "shisa-v2.1c-lfm2-350m-sft3-tlonly"

# GBS = 128 / 8 GPU / 16 MBS / 1 GAS
gradient_accumulation_steps: 1
micro_batch_size: 16
num_epochs: 1
optimizer: adamw_torch_4bit
lr_scheduler: cosine
learning_rate: 6e-5 # 4.78 @ GBS=128

train_on_inputs: false
group_by_length: false
bf16: true
tf32: false

gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
logging_steps: 1

warmup_ratio: 0.03
saves_per_epoch: 1

deepspeed: zero3_bf16.json
weight_decay: 1e-4

data/outputs/shisa-v2.1c-lfm2-350m-sft3-tlonly

This model was trained from scratch on the chotto-20251010.sft-tlonly.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: 6e-05
• train_batch_size: 16
• eval_batch_size: 16
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• total_train_batch_size: 128
• total_eval_batch_size: 128
• optimizer: Use OptimizerNames.ADAMW_TORCH_4BIT 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: 5
• training_steps: 187

Training results

Framework versions

• Transformers 4.57.0
• Pytorch 2.8.0+rocm6.4
• Datasets 4.1.1
• Tokenizers 0.22.1