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/dataset-scripts/preprocess_dataset/preprocess_diff_mask_chat.py
# (from felixwangg/stage_2_secure; token-mode diff, skip_indent, ctx=0).
# Columns: input_ids, attention_mask, labels, diff_mask.
# Labels are already -100 for non-assistant tokens; axolotl keeps them as-is.
datasets:
  - path: felixwangg/stage_2_secure_token_diff_mask_skip_indent_ctx0_chat
    type: pretokenized
    split: train
test_datasets:
  - path: felixwangg/stage_2_secure_token_diff_mask_skip_indent_ctx0_chat
    type: pretokenized
    split: validation
dataset_prepared_path: /home/tkwang/links/scratch/SecSteer-v2/axolotl-datasets/lora/Qwen2.5-Coder-7B/stage_2_secure_token_diff_mask_skip_indent_ctx0_chat
val_set_size: 0
output_dir: /home/tkwang/links/scratch/SecSteer-v2/axolotl-outputs/lora/Qwen2.5-Coder-7B-stage2-secure-token-diff-ctx0
sequence_len: 4096
sample_packing: false
eval_sample_packing: false
pad_to_sequence_len: true

adapter: lora
lora_model_dir: /home/tkwang/links/scratch/SecSteer-v2/axolotl-outputs/lora/Qwen2.5-Coder-7B-stage1-combined/checkpoint-6
lora_r: 16
lora_alpha: 16
lora_dropout: 0.05
lora_target_linear: true
merge_lora: false

wandb_project: diff-mask-stage1-2-ctx-0
wandb_entity: wtkuan
wandb_watch: "false"
wandb_name: Qwen2.5-Coder-7B-stage2-secure-token-diff-ctx0
wandb_log_model: "false"

gradient_accumulation_steps: 4
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: 2
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: 0.5

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

home/tkwang/links/scratch/SecSteer-v2/axolotl-outputs/lora/Qwen2.5-Coder-7B-stage2-secure-token-diff-ctx0

This model is a fine-tuned version of Qwen/Qwen2.5-Coder-7B-Instruct on the felixwangg/stage_2_secure_token_diff_mask_skip_indent_ctx0_chat dataset. It achieves the following results on the evaluation set:

Loss: 0.7459
Ppl: 2.1082
Memory/max Active (gib): 42.7
Memory/max Allocated (gib): 42.7
Memory/device Reserved (gib): 62.92

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: 4
gradient_accumulation_steps: 4
total_train_batch_size: 64
total_eval_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: 11
training_steps: 115

Training results

Training Loss	Epoch	Step	Validation Loss	Ppl	Active (gib)	Allocated (gib)	Reserved (gib)
No log	0	0	0.8545	2.3503	42.36	42.36	52.88
3.4189	0.2609	15	0.8128	2.2541	42.7	42.7	60.61
3.1757	0.5217	30	0.7668	2.1528	42.7	42.7	62.92
3.0517	0.7826	45	0.7548	2.1272	42.7	42.7	62.92
3.203	1.0348	60	0.7496	2.1161	42.7	42.7	62.92
2.9977	1.2957	75	0.7472	2.1111	42.7	42.7	62.92
2.9272	1.5565	90	0.7461	2.1088	42.7	42.7	62.92
2.8796	1.8174	105	0.7459	2.1082	42.7	42.7	62.92

Framework versions

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

felixwangg/Qwen2.5-Coder-7B-stage2-secure-token-diff-ctx0

home/tkwang/links/scratch/SecSteer-v2/axolotl-outputs/lora/Qwen2.5-Coder-7B-stage2-secure-token-diff-ctx0

Model description

Intended uses & limitations

Training and evaluation data

Training procedure

Training hyperparameters

Training results

Framework versions

felixwangg/
Qwen2.5-Coder-7B-stage2-secure-token-diff-ctx0