Qwen2.5-14B-style-SFT

<!-- This model card has been generated automatically according to the information the Trainer had access to. You should probably proofread and complete it, then remove this comment. -->

<img src="https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>

<details><summary>See axolotl config</summary>

axolotl version: 0.5.0

# =============================================================================
# PHASE 2 : SFT Réécriture sur modèle stylé (merged Phase 1)
# =============================================================================

base_model: philipperen55/Qwen2.5-14B-style-MERGED-v2
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer
trust_remote_code: true
subfolder: merged

load_in_8bit: false
load_in_4bit: false

# =============================================================================
# DATASETS : Paires SFT + Replay Buffer
# =============================================================================
datasets:
  # Dataset principal : paires de réécriture (format Alpaca)
  - path: philipperen55/phase2_paires
    data_files: phase2_paires.jsonl
    type: alpaca
    
  # Replay buffer : textes bruts Phase 1 (format Completion)
  - path: philipperen55/phase1_CLM
    data_files: phase1_CLM.jsonl
    type: completion
    field: text

dataset_prepared_path: /workspace/prepared_data
val_set_size: 0.0
output_dir: /workspace/output

# =============================================================================
# SÉQUENCE ET PACKING
# =============================================================================
sequence_len: 2048
sample_packing: true
pad_to_sequence_len: true
train_on_inputs: false

# =============================================================================
# LoRA - CONSERVATEUR (Rank 32 / Alpha 32)
# =============================================================================
adapter: lora
lora_r: 32
lora_alpha: 32
lora_dropout: 0.05
lora_target_modules:
  - q_proj
  - k_proj
  - v_proj
  - o_proj
  - gate_proj
  - up_proj
  - down_proj

# =============================================================================
# TRAINING
# =============================================================================
gradient_accumulation_steps: 2
micro_batch_size: 8
num_epochs: 2
learning_rate: 5e-6
lr_scheduler: cosine
warmup_ratio: 0.05
optimizer: adamw_torch

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

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

# =============================================================================
# LOGGING ET SAUVEGARDES (toutes les ~0.5 epoch)
# =============================================================================
# Calcul : ~8575 exemples, packing → ~4000 séquences
# Batch effectif = 8 × 2 = 16 → ~250 steps/epoch
# 0.5 epoch ≈ 125 steps
logging_steps: 10
save_steps: 125
save_total_limit: 6

# =============================================================================
# WEIGHTS & BIASES
# =============================================================================
wandb_project: Qwen2.5-14B-SFT-style-phase2

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

</details><br>

Qwen2.5-14B-style-SFT

This model is a fine-tuned version of philipperen55/Qwen2.5-14B-style-MERGED-v2 on the None 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: 5e-06
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 16
• 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: 28
• num_epochs: 2

Training results

Framework versions

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