You can deploy this model and start earning money today!

philipperen55/
Qwen2.5-14B-dataset41SFT

Name: Qwen2.5-14B-dataset41SFT
Availability: InStock

generated_from_trainer

philipperen55/Qwen2.5-14B-style-MERGED-v3

adapter

Views

No views yet

See axolotl config

axolotl version: 0.5.0

yaml

# =============================================================================
# PHASE 2 V3 : SFT Réécriture - FORMAT COMPLETION (sans template Alpaca)
# =============================================================================

base_model: philipperen55/Qwen2.5-14B-style-MERGED-v3
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer
trust_remote_code: true

load_in_8bit: false
load_in_4bit: false

# =============================================================================
# DATASET : Format completion pré-formaté (pas de template ajouté par Axolotl)
# =============================================================================
datasets:
  - path: philipperen55/dataset41SFT
    data_files: dataset41SFT.jsonl
    type: completion
    field: text

dataset_prepared_path: /workspace/prepared_data
val_set_size: 0.01
output_dir: /workspace/output

# =============================================================================
# SÉQUENCE - SANS PACKING
# =============================================================================
sequence_len: 2048
sample_packing: false
pad_to_sequence_len: true

# =============================================================================
# TOKEN EOS EXPLICITE
# =============================================================================
special_tokens:
  eos_token: "<|endoftext|>"
  pad_token: "<|endoftext|>"

# =============================================================================
# LoRA
# =============================================================================
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

# =============================================================================
# TRAINING - 40K PAIRES, 1 EPOCHS, LR 1e-5
# =============================================================================
gradient_accumulation_steps: 8
micro_batch_size: 4
num_epochs: 1
learning_rate: 1e-5
lr_scheduler: cosine
warmup_ratio: 0.06
optimizer: adamw_torch
weight_decay: 0.01
max_grad_norm: 1.0

# =============================================================================
# PRÉCISION
# =============================================================================
bf16: true
fp16: false
tf32: true

# =============================================================================
# OPTIMISATIONS
# =============================================================================
flash_attention: false
gradient_checkpointing: true

# --- Reproductibilité ---
seed: 42

# =============================================================================
# LOGGING ET SAUVEGARDES
# =============================================================================
logging_steps: 10
save_steps: 200
eval_strategy: steps 
eval_steps: 200
save_total_limit: 8

# =============================================================================
# WEIGHTS & BIASES
# =============================================================================
wandb_project: Qwen2.5-14B-dataset41SFT

# =============================================================================
# HUGGING FACE HUB
# =============================================================================
hub_model_id: philipperen55/Qwen2.5-14B-dataset41SFT
push_to_hub: true
hub_strategy: every_save

Qwen2.5-14B-dataset41SFT

This model is a fine-tuned version of philipperen55/Qwen2.5-14B-style-MERGED-v3 on the None dataset. It achieves the following results on the evaluation set:

Loss: 1.6398

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: 4
eval_batch_size: 4
seed: 42
gradient_accumulation_steps: 8
total_train_batch_size: 32
optimizer: Use OptimizerNames.ADAMW_TORCH 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: 74
num_epochs: 1

Training results

Training Loss	Epoch	Step	Validation Loss
No log	0.0008	1	1.8784
1.6592	0.1616	200	1.6617
1.6717	0.3232	400	1.6493
1.6543	0.4848	600	1.6442
1.6484	0.6465	800	1.6413
1.6569	0.8081	1000	1.6402
1.6387	0.9697	1200	1.6398

Framework versions

PEFT 0.13.2
Transformers 4.46.1
Pytorch 2.3.1+cu121
Datasets 3.0.1
Tokenizers 0.20.3