shisa-ai/gemma3.paradox.v2.121

See axolotl config

axolotl version: 0.8.0.dev0

base_model: google/gemma-3-27b-it
# optionally might have model_type or tokenizer_type
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer
# Automatically upload checkpoint and final model to HF
# hub_model_id: username/custom_model_name

load_in_8bit: false
load_in_4bit: false
strict: false

# huggingface repo
chat_template: gemma3
datasets:
  - path: shisa-ai/paradox_test_set_200k_sharegpt-v2
    type: chat_template
    field_messages: conversations
    message_property_mappings:
      role: role
      content: content
    split: train[:25%]
    
val_set_size: 0.0
output_dir: ./outputs/ablation-121-gemma3.paradox.v2


sequence_len: 8196
sample_packing: true
eval_sample_packing: false
pad_to_sequence_len: true

wandb_project:
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:


gradient_accumulation_steps: 2
micro_batch_size: 1
num_epochs: 1
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 6.53e-6

train_on_inputs: false
group_by_length: false
bf16: auto
fp16:
tf32: true

gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: false
eager_attention: true

warmup_ratio: 0.1
evals_per_epoch:
eval_table_size:
eval_max_new_tokens: 128
saves_per_epoch: 1
debug:
deepspeed: zero3_bf16.json
weight_decay: 0.05
fsdp:
fsdp_config:
special_tokens:
saves_per_epoch: 0
save_total_limit: 1 # Only store a single checkpoint

outputs/ablation-121-gemma3.paradox.v2

This model is a fine-tuned version of google/gemma-3-27b-it on the shisa-ai/paradox_test_set_200k_sharegpt-v2 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: 6.53e-06
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• distributed_type: multi-GPU
• num_devices: 48
• gradient_accumulation_steps: 2
• total_train_batch_size: 96
• total_eval_batch_size: 48
• optimizer: Use OptimizerNames.ADAMW_BNB 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: 19
• num_epochs: 1.0

Training results

Framework versions

• Transformers 4.50.0
• Pytorch 2.6.0+cu124
• Datasets 3.4.1
• Tokenizers 0.21.1