base_model: swiss-ai/Apertus-8B-Instruct-2509
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer
trust_remote_code: true

load_in_4bit: true
bnb_4bit_quant_type: nf4
bnb_4bit_compute_dtype: bfloat16
bnb_4bit_use_double_quant: true

datasets:
  - path: Jackrong/DeepSeek-V4-Distill-8000x
    type:
      field_instruction: input
      field_output: output

      format: |
        ### Instruction:
        {instruction}

        ### Response:

      no_input_format: |
        ### Instruction:
        {instruction}

        ### Response:
  
#dataset_prepared_path: /data/prepared/apertus8b_deepseekv4

output_dir: /outputs/apertus-8b-deepseekv4-qlora

val_set_size: 0.05

sequence_len: 8192
sample_packing: true
pad_to_sequence_len: true

adapter: qlora

lora_r: 32
lora_alpha: 64
lora_dropout: 0.05
lora_target_linear: true

micro_batch_size: 2
gradient_accumulation_steps: 8

num_epochs: 3

optimizer: paged_adamw_8bit

learning_rate: 0.00015
lr_scheduler: cosine
warmup_ratio: 0.03
weight_decay: 0.0

bf16: true
tf32: true

gradient_checkpointing: true

logging_steps: 10
eval_steps: 100
save_steps: 100
save_total_limit: 3

#############################################
# HUGGING FACE HUB
#############################################

hub_model_id: laurent-maille/apertus-8b-deepseekv4-fr
hub_strategy: every_save


hf_use_auth_token: true

hub_private_repo: true

save_safetensors: true

#############################################
# OPTIONAL
#############################################

#wandb_project: apertus-8b-ft
#wandb_name: apertus-8b-deepseekv4

special_tokens:
  pad_token: "<|endoftext|>"

apertus-8b-deepseekv4-fr

This model is a fine-tuned version of swiss-ai/Apertus-8B-Instruct-2509 on the Jackrong/DeepSeek-V4-Distill-8000x dataset. It achieves the following results on the evaluation set:

Loss: 1.1878
Ppl: 3.2797
Memory/max Active (gib): 26.08
Memory/max Allocated (gib): 26.08
Memory/device Reserved (gib): 43.59

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: 0.00015
train_batch_size: 2
eval_batch_size: 2
seed: 42
gradient_accumulation_steps: 8
total_train_batch_size: 16
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: cosine
lr_scheduler_warmup_steps: 10
training_steps: 360

Training results

Training Loss	Epoch	Step	Validation Loss	Ppl	Active (gib)	Allocated (gib)	Reserved (gib)
No log	0	0	4.0401	56.8296	26.06	26.06	27.67
1.2896	0.8282	100	1.2931	3.6441	26.08	26.08	43.39
1.2472	1.6625	200	1.2097	3.3524	26.08	26.08	43.84
1.2004	2.4886	300	1.1893	3.2847	26.08	26.08	43.84
1.1840	2.9855	360	1.1878	3.2797	26.08	26.08	43.59

Framework versions

PEFT 0.19.1
Transformers 5.5.4
Pytorch 2.10.0+cu128
Datasets 4.8.5
Tokenizers 0.22.2