See axolotl config

axolotl version: 0.10.0

base_model: meta-llama/Llama-3.2-3B-Instruct

load_in_8bit: true
load_in_4bit: false

datasets:
  - path: ./data/train_openai_response_transformed.jsonl
    type: chat_template

    field_messages: messages
    message_property_mappings:
      role: role
      content: content

val_file: ./data/val_openai_response_transformed.jsonl
val_set_size: 0.0
output_dir: ./outputs/cf-llm-finetune-llama-3.2-3b-lora

adapter: lora
lora_model_dir:

sequence_len: 4096
sample_packing: false
eval_sample_packing: true
pad_to_sequence_len: true

lora_r: 32
lora_alpha: 16
lora_dropout: 0.05
lora_target_linear: true

wandb_project:
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 4
micro_batch_size: 2
num_epochs: 4

optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.0002

bf16: auto
tf32: false

gradient_checkpointing: true
resume_from_checkpoint:
logging_steps: 1
flash_attention: false

warmup_steps: 10
evals_per_epoch: 4
saves_per_epoch: 1
weight_decay: 0.0
special_tokens:
  pad_token: "<|end_of_text|>"

Llama-3.2-3B-Instruct-PEFT-code-generation

This model is a fine tuned meta-llama/Llama-3.2-3B-Instruct on a synthetic dataset of C++ → Python code translations from Codeforces.

📦 GitHub repo: DemoVersion/cf-llm-finetune

📑 Dataset Creation DATASET.md

📑 Training TRAIN.md

📚 Dataset on Hugging Face: demoversion/cf-cpp-to-python-code-generation

For dataset generation, training, and inference check the Github repo.

📚 Main medium article: Toward fine-tuning Llama 3.2 using PEFT for Code Generation

📚 Medium article for inference with GGUF format: How to inference with GGUF format

Model description

A lightweight LLaMA 3.2 model fine-tuned for competitive programming code translation, from ICPC-style C++ to Python using LoRA adapters.

Intended uses & limitations

Use for:

• Translating competitive programming C++ solutions to Python
• Code understanding in educational or automation tools

Limitations:

• Not general-purpose code translation
• Python outputs are synthetically generated using GPT-4.1
• Focused only on ICPC-style problems

Training and evaluation data

Training and Evaluation data:
🧾 demoversion/cf-cpp-to-python-code-generation

Built from:

• open-r1/codeforces-submissions
• open-r1/codeforces

C++ submissions were filtered and paired with GPT-4.1-generated Python translations. Dataset split: 1,400 train / 300 val / 300 test. To underestand how the dataset was created check DATASET.md

Training procedure

• Adapter: LoRA (r=32, alpha=16, dropout=0.05)
• Optimizer: adamw_bnb_8bit
• LR: 2e-4, scheduler: cosine
• Batch size: 2 × 4 (grad accumulation) = total 8
• Training steps: 688
  Full config: TRAIN.md

Framework versions

• PEFT 0.15.2
• Transformers 4.52.3
• PyTorch 2.6.0+cu124
• Datasets 3.6.0
• Tokenizers 0.21.2