allura-forge/g12bsftep2

See axolotl config

axolotl version: 0.13.0.dev0

## model
base_model: ./model

## qlora COPE!!!
load_in_8bit: false
load_in_4bit: false #false
strict: false

# === Data Configuration ===
datasets:
  - path: allura-forge/expr-rp-sft-mix
    type: chat_template
    split: train
    field_messages: conversations
    message_field_role: from
    message_field_content: value

chat_template: jinja
chat_template_jinja: "{{ bos_token }}{% for message in messages %}{% if (message['role'] == 'assistant') %}{% set role = 'model' %}{% else %}{% set role = 'user' %}{% endif %}{{ '<start_of_turn>' + role + '\n' + message['content'] | trim + '<end_of_turn>\n' }}{% endfor %}{% if add_generation_prompt %}{{'<start_of_turn>model\n'}}{% endif %}"

shuffle_merged_datasets: true
dataset_prepared_path: dataset_prepareds
val_set_size: 0.0
output_dir: ./output

max_grad_norm: 0.1

## Liger + CCE
plugins:
  - axolotl.integrations.liger.LigerPlugin
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
liger_rope: true
liger_rms_norm: true
liger_layer_norm: true
liger_glu_activation: true
liger_fused_linear_cross_entropy: false
cut_cross_entropy: true

## CTX settings
sequence_len: 16384
sample_packing: true
eval_sample_packing: false
pad_to_sequence_len: true

## WandB
wandb_project: g12b-slopification
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

## hoe params
gradient_accumulation_steps: 2 # ???
micro_batch_size: 2
num_epochs: 2
lr_scheduler: rex
learning_rate: 2e-6
optimizer: adamw_torch_8bit # Options: "paged_ademamix_8bit", "adamw_bnb_8bit", "paged_adamw_8bit"

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

gradient_checkpointing: offload
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true
s2_attention:
special_tokens:
  eos_token: "<end_of_turn>"

warmup_steps: 25
saves_per_epoch: 4
debug:
weight_decay: 0.0
fsdp:
  - full_shard
  - auto_wrap
fsdp_config:
  fsdp_activation_checkpointing: true
  fsdp_limit_all_gathers: true
  fsdp_use_orig_params: false
  fsdp_cpu_ram_efficient_loading: true
  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
  fsdp_transformer_layer_cls_to_wrap: Gemma3DecoderLayer
  fsdp_state_dict_type: FULL_STATE_DICT
  fsdp_reshard_after_forward: true
  fsdp_version: 2

output

This model was trained from scratch on the allura-forge/expr-rp-sft-mix 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-06
• train_batch_size: 2
• eval_batch_size: 2
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 2
• total_train_batch_size: 32
• total_eval_batch_size: 16
• optimizer: Use OptimizerNames.ADAMW_TORCH_8BIT 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: 25
• training_steps: 538

Training results

Framework versions

• Transformers 4.57.1
• Pytorch 2.8.0+cu129
• Datasets 4.4.1
• Tokenizers 0.22.1