#base_model: ertghiu256/qwen3-4b-code-reasoning
base_model: Qwen/Qwen3-4B
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer

# QLoRA
adapter: lora
lora_r: 32
lora_alpha: 64
lora_dropout: 0.05
lora_target_modules: [q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj]

datasets:
  - path: laurent-maille/pcl-test-S27
    type: chat_template
    field_messages: messages
    conversation: chat
dataset_prepared_path: ./prepared/plc_sharegpt
output_dir: ./outputs/valuoty-indus-plc-4B

# SFT propre
train_on_inputs: false
mask_user_tokens: true

# Séquences
sequence_len: 2048
sample_packing: true
pad_to_sequence_len: false
group_by_length: true
flash_attn_impl: fa2   # "none" si FA2 non dispo

# Quantization / dtypes
load_in_4bit: true
load_in_8bit: false
bnb_4bit_quant_type: nf4
bnb_4bit_use_double_quant: true
bnb_4bit_compute_dtype: bfloat16
bf16: true
fp16: false

# Optim & training
optimizer: adamw_bnb_8bit
micro_batch_size: 10                 # 8–12 selon VRAM
gradient_accumulation_steps: 16      # effectif ≈ 192 (12×16)
gradient_checkpointing: true
max_grad_norm: 0.3
#resume_from_checkpoint: ./outputs/valuoty-indus-plc-4B/checkpoint-15/

learning_rate: 8.0e-5                # 5e-5 si tu veux lisser
weight_decay: 0.0
lr_scheduler_type: cosine
warmup_ratio: 0.03
num_train_epochs: 1
seed: 42

# Eval / logs / save
val_set_size: 2000
evaluation_strategy: steps
eval_steps: 800
logging_steps: 20
load_best_model_at_end: true
metric_for_best_model: loss
save_strategy: steps
save_steps: 1500
save_total_limit: 4
save_safetensors: true

outputs/valuoty-indus-plc-4B

This model is a fine-tuned version of Qwen/Qwen3-4B on the laurent-maille/pcl-test-S27 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: 8e-05
train_batch_size: 10
eval_batch_size: 10
seed: 42
gradient_accumulation_steps: 16
total_train_batch_size: 160
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
num_epochs: 1.0

Training results

Training Loss	Epoch	Step	Validation Loss
No log	0.0399	1	4.1037

Framework versions

PEFT 0.14.0
Transformers 4.55.4
Pytorch 2.5.1+cu124
Datasets 3.2.0
Tokenizers 0.21.0

laurent-maille/Valuoty-industry-plc-4B-V0.4Adap

outputs/valuoty-indus-plc-4B

Model description

Intended uses & limitations

Training and evaluation data

Training procedure

Training hyperparameters

Training results

Framework versions

laurent-maille/
Valuoty-industry-plc-4B-V0.4Adap