Qwen2.5-Coder-7B-sft-plus-alpha-0p1-line-diff-ctx0-v2

<!-- 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: 0.13.2

base_model: Qwen/Qwen2.5-Coder-7B-Instruct
model_type: Qwen2ForCausalLM
tokenizer_type: AutoTokenizer

load_in_8bit: false
load_in_4bit: false

# Pre-tokenized datasets produced by scripts/preprocess_diff_mask_chat.py
# (from felixwangg/prime_vul_plus_splitted which has a fixed validation split).
# Columns: input_ids, attention_mask, labels, diff_mask.
# Labels are already -100 for non-assistant tokens; axolotl keeps them as-is.
datasets:
  - path: felixwangg/prime_vul_plus_splitted_line_diff_mask_skip_indent_ctx0_chat_v2
    type: pretokenized
    split: train
test_datasets:
  - path: felixwangg/prime_vul_plus_splitted_line_diff_mask_skip_indent_ctx0_chat_v2
    type: pretokenized
    split: validation
dataset_prepared_path: /scratch/tkwang/SecSteer/axolotl-datasets/lora/Qwen2.5-Coder-7B/prime_vul_plus_splitted_line_diff_mask_skip_indent_ctx0_chat_v2_alpha_0p1
val_set_size: 0
output_dir: /scratch/tkwang/SecSteer/axolotl-outputs/lora/Qwen2.5-Coder-7B-sft-plus-alpha-0p1-line-diff-ctx0-v2
sequence_len: 4096
sample_packing: false
eval_sample_packing: false
pad_to_sequence_len: true

adapter: lora
lora_model_dir:
lora_r: 16
lora_alpha: 16
lora_dropout: 0.05
lora_target_linear: true
merge_lora: true

wandb_project: diff-mask-sft-primevul-ctx-0
wandb_entity: wtkuan
wandb_watch: "false"
wandb_name: Qwen2.5-Coder-7B-sft-plus-alpha-0p1-line-diff-ctx0-v2
wandb_log_model: "false"

gradient_accumulation_steps: 8
micro_batch_size: 4
optimizer: adamw_torch
lr_scheduler: cosine
learning_rate: 4e-05

bf16: true
tf32: false

gradient_checkpointing: true
resume_from_checkpoint:
logging_steps: 1
flash_attention: true

num_epochs: 1
warmup_ratio: 0.1
early_stopping_patience: 1000
eval_steps: 15
save_steps: 15
save_total_limit: 1000
load_best_model_at_end: true

weight_decay: 0.02
special_tokens:

# Diff-mask weighted loss: CE(logit_t, label_t) * (1 + alpha * diff_mask_{t+1})
# Security-sensitive tokens (diff_mask=1) get weight (1 + diff_mask_alpha).
# Requires PYTHONPATH to include the repo root so diff_mask_trainer is importable.
diff_mask_alpha: 0.1

plugins:
  - diff_mask_trainer.plugin.DiffMaskPlugin
  # - sec_bench_callback.SecBenchPlugin

</details><br>

scratch/tkwang/SecSteer/axolotl-outputs/lora/Qwen2.5-Coder-7B-sft-plus-alpha-0p1-line-diff-ctx0-v2

This model is a fine-tuned version of Qwen/Qwen2.5-Coder-7B-Instruct on the felixwangg/prime_vul_plus_splitted_line_diff_mask_skip_indent_ctx0_chat_v2 dataset. It achieves the following results on the evaluation set:

• Loss: 0.7546
• Ppl: 2.1268
• Memory/max Active (gib): 42.7
• Memory/max Allocated (gib): 42.7
• Memory/device Reserved (gib): 61.19

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: 4e-05
• train_batch_size: 4
• eval_batch_size: 4
• seed: 42
• distributed_type: multi-GPU
• num_devices: 2
• gradient_accumulation_steps: 8
• total_train_batch_size: 64
• total_eval_batch_size: 8
• 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: 5
• training_steps: 57

Training results

Framework versions

• PEFT 0.18.1
• Transformers 4.57.6
• Pytorch 2.10.0+cu128
• Datasets 4.5.0
• Tokenizers 0.22.2