See axolotl config

axolotl version: 0.8.0.dev0

base_model: unsloth/phi-4
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer

load_in_8bit: false
load_in_4bit: false
strict: false

# User Liger
plugins:
  - axolotl.integrations.liger.LigerPlugin
liger_rope: true
liger_rms_norm: true
liger_glu_activation: true
liger_fused_linear_cross_entropy: true

chat_template: tokenizer_default
datasets:
  - path: shisa-ai/shisa-v2-best-of-n-athenev2-tulu70b-llama33-only-no-sysprompt
    type: chat_template
    field_messages: conversations
    message_field_role: from
    message_field_content: value
  - path: shisa-ai/shisa-v2-roleplaying-sft
    type: chat_template
    field_messages: conversations
    message_property_mappings:
      role: role
      content: content
    roles:
      system:
        - system
      assistant:
        - gpt
        - model
        - assistant
      user:
        - human
        - user
    roles_to_train: ["assistant"]
  - path: shisa-ai/translation_expanded_master_set_filtered
    split: train[:25%]
    type: chat_template
    field_messages: conversations
    message_property_mappings:
      role: role
      content: content
    roles:
      system:
        - system
      assistant:
        - gpt
        - model
        - assistant
      user:
        - human
        - user
    roles_to_train: ["assistant"]
  - path: shisa-ai/rewild-set
    split: train[:5%]
    type: chat_template
    field_messages: conversations
    message_property_mappings:
      role: role
      content: content
    roles:
      system:
        - system
      assistant:
        - gpt
        - model
        - assistant
      user:
        - human
        - user
    roles_to_train: ["assistant"]
  - path: shisa-ai/magpie-ultra-set
    split: train[:8%]
    type: chat_template
    field_messages: conversations
    message_property_mappings:
      role: role
      content: content
    roles:
      system:
        - system
      assistant:
        - gpt
        - model
        - assistant
      user:
        - human
        - user
    roles_to_train: ["assistant"]
  - path: shisa-ai/magpie-advanced-questions-set
    split: train[:8%]
    type: chat_template
    field_messages: conversations
    message_property_mappings:
      role: role
      content: content
    roles:
      system:
        - system
      assistant:
        - gpt
        - model
        - assistant
      user:
        - human
        - user
    roles_to_train: ["assistant"]
  - path: shisa-ai/japan-magpie-set
    split: train
    type: chat_template
    field_messages: conversations
    message_property_mappings:
      role: role
      content: content
    roles:
      system:
        - system
      assistant:
        - gpt
        - model
        - assistant
      user:
        - human
        - user
    roles_to_train: ["assistant"]

dataset_prepared_path: last_run_prepared
val_set_size: 0.05
output_dir: ./outputs/ablation-114-bestofn.atlo.rp.tlx.newmix2.unphi-shisa-v2-unphi-4-14b

sequence_len: 8192
sample_packing: true
pad_to_sequence_len: true

# marginal difference
neftune_noise_alpha: 5

use_wandb: true
wandb_project: shisa-v2
wandb_entity: augmxnt
wandb_name: ablation-114-bestofn.atlo.rp.tlx.newmix2.unphi-shisa-v2-unphi-4-14b

gradient_accumulation_steps: 1
micro_batch_size: 4
num_epochs: 3
optimizer: paged_adamw_8bit
lr_scheduler: linear
learning_rate: 5e-6

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

gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
early_stopping_patience:
resume_from_checkpoint:
logging_steps: 1
xformers_attention:
flash_attention: true

warmup_ratio: 0.1
evals_per_epoch: 4
eval_table_size:
saves_per_epoch: 0
save_total_limit: 1 # Only store a single checkpoint
debug:
deepspeed: zero3_bf16.json
weight_decay: 0.0001
fsdp:
fsdp_config:
special_tokens:
  # pad_token: "<|dummy_87|>"

outputs/ablation-114-bestofn.atlo.rp.tlx.newmix2.unphi-shisa-v2-unphi-4-14b

This model is a fine-tuned version of unsloth/phi-4 on the shisa-ai/shisa-v2-best-of-n-athenev2-tulu70b-llama33-only-no-sysprompt, the shisa-ai/shisa-v2-roleplaying-sft, the shisa-ai/translation_expanded_master_set_filtered, the shisa-ai/rewild-set, the shisa-ai/magpie-ultra-set, the shisa-ai/magpie-advanced-questions-set and the shisa-ai/japan-magpie-set datasets. It achieves the following results on the evaluation set:

• Loss: 0.4720

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: 5e-06
• train_batch_size: 4
• eval_batch_size: 4
• seed: 42
• distributed_type: multi-GPU
• num_devices: 16
• total_train_batch_size: 64
• total_eval_batch_size: 64
• optimizer: Use OptimizerNames.PAGED_ADAMW_8BIT with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 296
• num_epochs: 3.0

Training results

Training Loss	Epoch	Step	Validation Loss
0.7211	0.0010	1	0.7194
0.542	0.2508	248	0.5293
0.4985	0.5015	496	0.5020
0.4755	0.7523	744	0.4898
0.4368	1.0030	992	0.4833
0.4605	1.2538	1240	0.4796
0.4388	1.5046	1488	0.4758
0.4494	1.7553	1736	0.4723
0.3682	2.0061	1984	0.4753
0.4104	2.2568	2232	0.4742
0.3903	2.5076	2480	0.4732
0.3818	2.7583	2728	0.4720

Framework versions

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