# !pip install transformers==4.55.4
# !pip install --no-deps trl==0.22.2
# !pip install --no-build-isolation mamba_ssm==2.2.5
# !pip install --no-build-isolation causal_conv1d==1.5.2
# === Model Configuration ===
base_model: apertus/trained-instruct-attn
load_in_8bit: false
load_in_4bit: false

# === HF Configuration === 
#hub_model_id: ToastyPigeon/muse-marvin-32k-lora-2
#hub_strategy: "every_save"
output_dir: apertus/trained-again-instruct-o-down

# === Wandb Tracking ===
wandb_project: ApertusTests
# wandb_entity: [WANDB_ENTITY]
wandb_name: trained-again-instruct-o-down

# === Training Setup ===
num_epochs: 1
micro_batch_size: 1
gradient_accumulation_steps: 4
sequence_len: 4096
#sequence_parallel_degree: 2
#heads_k_stride: 1
sample_packing: true
#pad_to_sequence_len: true
#temperature: 0.7
#max_steps: 10
# === Evaluation ===
val_set_size: 0.025
evals_per_epoch: 10
#eval_steps: 20
#max_steps: 60
#eval_table_size:
eval_max_new_tokens: 128
#eval_sample_packing: true
#eval_strategy: "no"

# === LoRA Configuration ===
adapter:
lora_model_dir:
lora_r:
lora_alpha:
lora_dropout:
lora_target_linear:
lora_target_modules:
#  - up_proj
#  - down_proj
#  - gate_proj
#  - q_proj
#  - v_proj
#  - k_proj
#  - o_proj
#  - input_layernorm
#  - post_attention_layernorm
#  - embed_tokens
#  - lm_head

lora_fan_in_fan_out:
#peft_use_rslora: true
lora_modules_to_save:
#  - embed_tokens
#  - lm_head
#fix_untrained_tokens: true
#lora_mlp_kernel: true
#lora_qkv_kernel: true
#lora_o_kernel: true
unfrozen_parameters:
  - model.layers.[0-9]+.self_attn.o_proj
  - model.layers.[0-9]+.mlp.down_proj
# === Hyperparameter Configuration ===
#optimizer: apollo_adamw_layerwise
#warmup_steps: 0
warmup_ratio: 0.025
optimizer: adamw_torch_fused
#optimizer: paged_adamw_8bit
#optim_args:
#  enable_stochastic_rounding: true
#  enable_cautious: true
#  enable_8bit: true
# Apollo-mini configuration:
#optim_args: "proj=random,rank=128,scale=128.0,scale_type=tensor,update_proj_gap=100"
# Regular Apollo configuration:
# optim_args: 
#optim_target_modules: all_linear
learning_rate: 1e-5
lr_scheduler: cosine
#cosine_min_lr_ratio: 0.2
#lr_scheduler: cosine_with_min_lr
#lr_scheduler_kwargs:
#  cosine_min_lr: 1e-6
weight_decay: 0.01
max_grad_norm: 1.0
#warmup_steps: 0
#warmup_ratio: 0.025


# === Data Configuration ===
#
#chat_template: jinja
chat_template: chatml
special_tokens:
#  eos_token: "<|im_end|>"
#  eos_token: "</s>"
#tokenizer_use_mistral_common: true
shuffle_merged_datasets: true
datasets:
#  - path: grimulkan/LimaRP-augmented
#    type: chat_template
#    field_messages: conversations
#    message_property_mappings:
#      role: from
#      content: value
#  - path: allenai/tulu-3-sft-personas-instruction-following
#    type: chat_template
#    split: train[:10%]
#  - path: ToastyPigeon/mixed-medical-reasoning-formatted
#    type: chat_template
#    data_files: mixed-medical-thinking.json
#    split: train[:10%]
#  - path: ToastyPigeon/steve-and-marvin
#    type: completion
#    data_files: marvin.json
#  - path: ToastyPigeon/kimi-stories-completion
#    type: completion
#  - path: ToastyPigeon/new-story-dataset
 #   type: customcompletion-regex
#    type: completion
#    data_files: new-story-dataset-v2.json
#  - path: allura-org/fujin-instruct-v2
#    type: customchatml-regex
#    type: chat_template
#    field_messages: conversations
#    message_property_mappings:
#      role: from
#      content: value
#  - path: ToastyPigeon/some-rp-extended
 #   type: customchatml-regex
#    type: chat_template
#    field_messages: conversations
#    message_property_mappings:
#      role: from
#      content: value
#    roles_to_train: ["user","assistant"]
  - path: allura-forge/koto-instruct-sft
#    type: customchatml-regex
    type: chat_template
    split: train[50%:]
    field_messages: conversations
    message_property_mappings:
      role: from
      content: value
#  - path: ToastyPigeon/SpringDragon
#    type: customcompletion-regex
#    type: completion
#    split: train
#  - path: ToastyPigeon/some-erotica
#    type: customcompletion-regex
#    type: completion
#    split: train[:10%]
#  - path: ToastyPigeon/tulu-mini
#    type: chat_template
dataset_prepared_path: last_run_prepared


# === Plugins ===
plugins:
  - axolotl.integrations.liger.LigerPlugin
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin

# === Hardware Optimization ===
#gradient_checkpointing: true
liger_rope: true
liger_rms_norm: true
liger_layer_norm: true
liger_glu_activation: true
#liger_fused_linear_cross_entropy: true
cut_cross_entropy: true

#deepspeed: ../axolotl/deepspeed_configs/zero2.json

# === FSDP Config === 
fsdp:
  - full_shard
  - auto_wrap
fsdp_config:
  fsdp_limit_all_gathers: true
  fsdp_sync_module_states: true
  fsdp_offload_params: true
  fsdp_activation_checkpointing: true
  fsdp_use_orig_params: true
  fsdp_cpu_ram_efficient_loading: true
  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
  fsdp_transformer_layer_cls_to_wrap: ApertusDecoderLayer
  fsdp_state_dict_type: FULL_STATE_DICT
  fsdp_sharding_strategy: FULL_SHARD

# === Checkpointing ===
#save_steps: 10
saves_per_epoch: 1
save_total_limit: 1

# === Advanced Settings ===
bf16: auto
flash_attention: true
train_on_inputs: false
group_by_length: false
save_safetensors: true
logging_steps: 1
gc_steps: 10
seed: 69

apertus/trained-again-instruct-o-down

This model was trained from scratch on the allura-forge/koto-instruct-sft dataset. It achieves the following results on the evaluation set:

Loss: 0.9467
Memory/max Active (gib): 5.15
Memory/max Allocated (gib): 5.15
Memory/device Reserved (gib): 6.41

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: 1e-05
train_batch_size: 1
eval_batch_size: 1
seed: 69
distributed_type: multi-GPU
num_devices: 2
gradient_accumulation_steps: 4
total_train_batch_size: 8
total_eval_batch_size: 2
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: 12
training_steps: 516

Training results

Training Loss	Epoch	Step	Validation Loss	Active (gib)	Allocated (gib)	Reserved (gib)
No log	0	0	1.0191	6.25	5.19	6.43
0.8751	0.1008	52	0.9954	5.15	5.15	6.41
1.0313	0.2016	104	0.9796	5.15	5.15	6.41
1.0144	0.3023	156	0.9677	5.15	5.15	6.41
1.0103	0.4031	208	0.9606	5.15	5.15	6.41
0.862	0.5039	260	0.9553	5.15	5.15	6.41
0.9892	0.6047	312	0.9512	5.15	5.15	6.41
1.0593	0.7054	364	0.9488	5.15	5.15	6.41
0.9527	0.8062	416	0.9474	5.15	5.15	6.41
0.8602	0.9070	468	0.9467	5.15	5.15	6.41

Framework versions

Transformers 4.56.1
Pytorch 2.7.1+cu126
Datasets 4.0.0
Tokenizers 0.22.1

allura-forge/birdpertus-instruct

apertus/trained-again-instruct-o-down

Model description

Intended uses & limitations

Training and evaluation data

Training procedure

Training hyperparameters

Training results

Framework versions

allura-forge/
birdpertus-instruct