base_model: NewEden/MistralAI-Nemo-Instruct-ChatML
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer

load_in_8bit: false
load_in_4bit: false
strict: false

datasets:
  - path: hardlyworking/HardlyRP
    type: chat_template
    chat_template: chatml
    roles_to_train: ["gpt"]
    field_messages: conversations
    message_field_role: from
    message_field_content: value
    train_on_eos: turn
  - path: jeiku/Writing
    type: completion
    field: text

shuffle_merged_datasets: true
dataset_prepared_path: dataset_preparedss
val_set_size: 0.0025
output_dir: 12b-out-0001-max_grad_norm

hub_model_id: hardlyworking/Sapphire-12B
hub_strategy: "all_checkpoints"
push_dataset_to_hub:
hf_use_auth_token: true

plugins:
  - axolotl.integrations.liger.LigerPlugin
liger_rope: true
liger_rms_norm: true
liger_layer_norm: true
liger_glu_activation: true
liger_fused_linear_cross_entropy: true

sequence_len: 8192
sample_packing: true
eval_sample_packing: false
pad_to_sequence_len: true

max_grad_norm: 0.001

wandb_project: Sapphire
wandb_entity:
wandb_watch:
wandb_name: Sapphire
wandb_log_model:

evals_per_epoch: 8
eval_table_size:
eval_max_new_tokens: 128

gradient_accumulation_steps: 8
micro_batch_size: 2
num_epochs: 2
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 2e-6

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

gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true
s2_attention:

warmup_ratio: 0.05
saves_per_epoch: 1
debug:
weight_decay: 0.0001
fsdp:
fsdp_config:
special_tokens:
   pad_token: <pad>

Sapphire-12B

This model is a fine-tuned version of NewEden/MistralAI-Nemo-Instruct-ChatML on the hardlyworking/HardlyRP and the jeiku/Writing datasets. It achieves the following results on the evaluation set:

Loss: 1.6799

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
gradient_accumulation_steps: 8
total_train_batch_size: 16
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: 30
num_epochs: 2.0

Training results

Training Loss	Epoch	Step	Validation Loss
1.8932	0.0033	1	1.9155
1.7729	0.1262	38	1.7802
1.7163	0.2525	76	1.7111
1.6484	0.3787	114	1.6970
1.7006	0.5050	152	1.6907
1.7276	0.6312	190	1.6874
1.7042	0.7575	228	1.6847
1.5575	0.8837	266	1.6825
1.5451	1.0100	304	1.6816
1.6592	1.1362	342	1.6807
1.7344	1.2625	380	1.6805
1.6953	1.3887	418	1.6798
1.5799	1.5150	456	1.6799
1.5241	1.6412	494	1.6799
1.548	1.7674	532	1.6797
1.6254	1.8937	570	1.6799

Framework versions

Transformers 4.51.0
Pytorch 2.6.0+cu124
Datasets 3.5.0
Tokenizers 0.21.1