============================================================
  SecureReview SFT Training
  Model : unsloth/Qwen2.5-7B-Instruct-bnb-4bit
  Task  : migration_review
  Epochs: 3
============================================================

[1/6] Checking environment connection...
  Health: {'status': 'healthy'}

[2/6] Loading model...
🦥 Unsloth: Will patch your computer to enable 2x faster free finetuning.
Unsloth: Your Flash Attention 2 installation seems to be broken. Using Xformers instead. No performance changes will be seen.
🦥 Unsloth Zoo will now patch everything to make training faster!
==((====))==  Unsloth 2026.4.8: Fast Qwen2 patching. Transformers: 5.5.0.
   \\   /|    NVIDIA L40S. Num GPUs = 1. Max memory: 44.392 GB. Platform: Linux.
O^O/ \_/ \    Torch: 2.10.0+cu128. CUDA: 8.9. CUDA Toolkit: 12.8. Triton: 3.6.0
\        /    Bfloat16 = TRUE. FA [Xformers = None. FA2 = False]
 "-____-"     Free license: http://github.com/unslothai/unsloth
Unsloth: Fast downloading is enabled - ignore downloading bars which are red colored!


model.safetensors:   0%|          | 0.00/5.55G [00:00<?, ?B/s]

model.safetensors:   0%|          | 0.00/5.55G [00:01<?, ?B/s]

model.safetensors:  20%|█▉        | 1.09G/5.55G [00:02<00:04, 942MB/s]

model.safetensors:  44%|████▍     | 2.46G/5.55G [00:03<00:02, 1.06GB/s]

model.safetensors:  66%|██████▌   | 3.67G/5.55G [00:04<00:01, 964MB/s] 

model.safetensors:  93%|█████████▎| 5.14G/5.55G [00:06<00:00, 942MB/s]
model.safetensors: 100%|██████████| 5.55G/5.55G [00:07<00:00, 790MB/s]


Loading weights:   0%|          | 0/339 [00:00<?, ?it/s]
Loading weights: 100%|██████████| 339/339 [00:00<00:00, 523.44it/s]


generation_config.json:   0%|          | 0.00/271 [00:00<?, ?B/s]
generation_config.json: 100%|██████████| 271/271 [00:00<00:00, 1.42MB/s]


tokenizer_config.json:   0%|          | 0.00/7.36k [00:00<?, ?B/s]
tokenizer_config.json: 100%|██████████| 7.36k/7.36k [00:00<00:00, 40.9MB/s]


vocab.json:   0%|          | 0.00/2.78M [00:00<?, ?B/s]
vocab.json: 100%|██████████| 2.78M/2.78M [00:00<00:00, 75.9MB/s]


merges.txt:   0%|          | 0.00/1.67M [00:00<?, ?B/s]
merges.txt: 100%|██████████| 1.67M/1.67M [00:00<00:00, 69.0MB/s]


tokenizer.json:   0%|          | 0.00/11.4M [00:00<?, ?B/s]
tokenizer.json: 100%|██████████| 11.4M/11.4M [00:00<00:00, 54.5MB/s]


added_tokens.json:   0%|          | 0.00/605 [00:00<?, ?B/s]
added_tokens.json: 100%|██████████| 605/605 [00:00<00:00, 3.87MB/s]


special_tokens_map.json:   0%|          | 0.00/614 [00:00<?, ?B/s]
special_tokens_map.json: 100%|██████████| 614/614 [00:00<00:00, 4.47MB/s]
unsloth/Qwen2.5-7B-Instruct-bnb-4bit does not have a padding token! Will use pad_token = <|PAD_TOKEN|>.
Unsloth 2026.4.8 patched 28 layers with 28 QKV layers, 28 O layers and 28 MLP layers.
trainable params: 40,370,176 || all params: 7,655,986,688 || trainable%: 0.5273

[3/6] Building SFT dataset from ground-truth findings...


ground_truth.json:   0%|          | 0.00/2.82k [00:00<?, ?B/s]
ground_truth.json: 100%|██████████| 2.82k/2.82k [00:00<00:00, 14.2MB/s]
  Loaded migration_002 (4 findings)


ground_truth.json:   0%|          | 0.00/3.64k [00:00<?, ?B/s]
ground_truth.json: 100%|██████████| 3.64k/3.64k [00:00<00:00, 26.0MB/s]
  Loaded migration_006 (6 findings)


ground_truth.json:   0%|          | 0.00/3.17k [00:00<?, ?B/s]
ground_truth.json: 100%|██████████| 3.17k/3.17k [00:00<00:00, 22.2MB/s]
  Loaded migration_007 (5 findings)


ground_truth.json:   0%|          | 0.00/3.39k [00:00<?, ?B/s]
ground_truth.json: 100%|██████████| 3.39k/3.39k [00:00<00:00, 25.9MB/s]
  Loaded migration_009 (5 findings)


ground_truth.json:   0%|          | 0.00/3.76k [00:00<?, ?B/s]
ground_truth.json: 100%|██████████| 3.76k/3.76k [00:00<00:00, 29.5MB/s]
  Loaded migration_012 (6 findings)


ground_truth.json:   0%|          | 0.00/4.35k [00:00<?, ?B/s]
ground_truth.json: 100%|██████████| 4.35k/4.35k [00:00<00:00, 28.9MB/s]
  Loaded migration_017 (6 findings)


ground_truth.json:   0%|          | 0.00/4.44k [00:00<?, ?B/s]
ground_truth.json: 100%|██████████| 4.44k/4.44k [00:00<00:00, 28.1MB/s]
  Loaded migration_018 (6 findings)


ground_truth.json:   0%|          | 0.00/3.98k [00:00<?, ?B/s]
ground_truth.json: 100%|██████████| 3.98k/3.98k [00:00<00:00, 28.0MB/s]
  Loaded migration_022 (6 findings)


ground_truth.json:   0%|          | 0.00/4.05k [00:00<?, ?B/s]
ground_truth.json: 100%|██████████| 4.05k/4.05k [00:00<00:00, 33.1MB/s]
  Loaded migration_023 (6 findings)


ground_truth.json:   0%|          | 0.00/4.49k [00:00<?, ?B/s]
ground_truth.json: 100%|██████████| 4.49k/4.49k [00:00<00:00, 29.5MB/s]
  Loaded migration_024 (6 findings)


ground_truth.json:   0%|          | 0.00/4.20k [00:00<?, ?B/s]
ground_truth.json: 100%|██████████| 4.20k/4.20k [00:00<00:00, 29.4MB/s]
  Loaded migration_025 (6 findings)


ground_truth.json:   0%|          | 0.00/4.00k [00:00<?, ?B/s]
ground_truth.json: 100%|██████████| 4.00k/4.00k [00:00<00:00, 28.7MB/s]
  Loaded migration_028 (6 findings)
  Dataset: 12 examples

[4/6] Baseline evaluation (before SFT)...
The attention mask is not set and cannot be inferred from input because pad token is same as eos token. As a consequence, you may observe unexpected behavior. Please pass your input's `attention_mask` to obtain reliable results.
Both `max_new_tokens` (=600) and `max_length`(=32768) seem to have been set. `max_new_tokens` will take precedence. Please refer to the documentation for more information. (https://huggingface.co/docs/transformers/main/en/main_classes/text_generation)
/root/.pyenv/versions/3.13.13/lib/python3.13/site-packages/transformers/modeling_attn_mask_utils.py:71: FutureWarning: The attention mask API under `transformers.modeling_attn_mask_utils` (`AttentionMaskConverter`) is deprecated and will be removed in Transformers v5.10. Please use the new API in `transformers.masking_utils`.
  warnings.warn(DEPRECATION_MESSAGE, FutureWarning)
/root/.pyenv/versions/3.13.13/lib/python3.13/site-packages/transformers/modeling_attn_mask_utils.py:281: FutureWarning: The attention mask API under `transformers.modeling_attn_mask_utils` (`AttentionMaskConverter`) is deprecated and will be removed in Transformers v5.10. Please use the new API in `transformers.masking_utils`.
  warnings.warn(DEPRECATION_MESSAGE, FutureWarning)
/root/.pyenv/versions/3.13.13/lib/python3.13/site-packages/transformers/modeling_attn_mask_utils.py:71: FutureWarning: The attention mask API under `transformers.modeling_attn_mask_utils` (`AttentionMaskConverter`) is deprecated and will be removed in Transformers v5.10. Please use the new API in `transformers.masking_utils`.
  warnings.warn(DEPRECATION_MESSAGE, FutureWarning)
/root/.pyenv/versions/3.13.13/lib/python3.13/site-packages/transformers/modeling_attn_mask_utils.py:281: FutureWarning: The attention mask API under `transformers.modeling_attn_mask_utils` (`AttentionMaskConverter`) is deprecated and will be removed in Transformers v5.10. Please use the new API in `transformers.masking_utils`.
  warnings.warn(DEPRECATION_MESSAGE, FutureWarning)
  [before] migration_002: 0.300
Both `max_new_tokens` (=600) and `max_length`(=32768) seem to have been set. `max_new_tokens` will take precedence. Please refer to the documentation for more information. (https://huggingface.co/docs/transformers/main/en/main_classes/text_generation)
/root/.pyenv/versions/3.13.13/lib/python3.13/site-packages/transformers/modeling_attn_mask_utils.py:71: FutureWarning: The attention mask API under `transformers.modeling_attn_mask_utils` (`AttentionMaskConverter`) is deprecated and will be removed in Transformers v5.10. Please use the new API in `transformers.masking_utils`.
  warnings.warn(DEPRECATION_MESSAGE, FutureWarning)
/root/.pyenv/versions/3.13.13/lib/python3.13/site-packages/transformers/modeling_attn_mask_utils.py:281: FutureWarning: The attention mask API under `transformers.modeling_attn_mask_utils` (`AttentionMaskConverter`) is deprecated and will be removed in Transformers v5.10. Please use the new API in `transformers.masking_utils`.
  warnings.warn(DEPRECATION_MESSAGE, FutureWarning)
  [before] migration_006: 0.520
Both `max_new_tokens` (=600) and `max_length`(=32768) seem to have been set. `max_new_tokens` will take precedence. Please refer to the documentation for more information. (https://huggingface.co/docs/transformers/main/en/main_classes/text_generation)
  [before] migration_007: 0.060
Both `max_new_tokens` (=600) and `max_length`(=32768) seem to have been set. `max_new_tokens` will take precedence. Please refer to the documentation for more information. (https://huggingface.co/docs/transformers/main/en/main_classes/text_generation)
  [before] migration_009: 0.260
Both `max_new_tokens` (=600) and `max_length`(=32768) seem to have been set. `max_new_tokens` will take precedence. Please refer to the documentation for more information. (https://huggingface.co/docs/transformers/main/en/main_classes/text_generation)
  [before] migration_012: 0.060
Both `max_new_tokens` (=600) and `max_length`(=32768) seem to have been set. `max_new_tokens` will take precedence. Please refer to the documentation for more information. (https://huggingface.co/docs/transformers/main/en/main_classes/text_generation)
  [before] migration_017: 0.060
Both `max_new_tokens` (=600) and `max_length`(=32768) seem to have been set. `max_new_tokens` will take precedence. Please refer to the documentation for more information. (https://huggingface.co/docs/transformers/main/en/main_classes/text_generation)
  [before] migration_018: 0.290
Both `max_new_tokens` (=600) and `max_length`(=32768) seem to have been set. `max_new_tokens` will take precedence. Please refer to the documentation for more information. (https://huggingface.co/docs/transformers/main/en/main_classes/text_generation)
  [before] migration_022: 0.060
Both `max_new_tokens` (=600) and `max_length`(=32768) seem to have been set. `max_new_tokens` will take precedence. Please refer to the documentation for more information. (https://huggingface.co/docs/transformers/main/en/main_classes/text_generation)
  [before] migration_023: 0.060
Both `max_new_tokens` (=600) and `max_length`(=32768) seem to have been set. `max_new_tokens` will take precedence. Please refer to the documentation for more information. (https://huggingface.co/docs/transformers/main/en/main_classes/text_generation)
  [before] migration_024: 0.280
Both `max_new_tokens` (=600) and `max_length`(=32768) seem to have been set. `max_new_tokens` will take precedence. Please refer to the documentation for more information. (https://huggingface.co/docs/transformers/main/en/main_classes/text_generation)
  [before] migration_025: 0.060
Both `max_new_tokens` (=600) and `max_length`(=32768) seem to have been set. `max_new_tokens` will take precedence. Please refer to the documentation for more information. (https://huggingface.co/docs/transformers/main/en/main_classes/text_generation)
  [before] migration_028: 0.030
  Baseline mean: 0.170

[5/6] SFT training...
warmup_ratio is deprecated and will be removed in v5.2. Use `warmup_steps` instead.
/app/unsloth_compiled_cache/UnslothSFTTrainer.py:915: UserWarning: Padding-free training is enabled, but the attention implementation is not set to 'flash_attention_2'. Padding-free training flattens batches into a single sequence, and 'flash_attention_2' is the only known attention mechanism that reliably supports this. Using other implementations may lead to unexpected behavior. To ensure compatibility, set `attn_implementation='flash_attention_2'` in the model configuration, or verify that your attention mechanism can handle flattened sequences.
  warnings.warn(


Unsloth: Tokenizing ["text"] (num_proc=12):   0%|          | 0/12 [00:00<?, ? examples/s]

Unsloth: Tokenizing ["text"] (num_proc=12):   8%|▊         | 1/12 [00:02<00:23,  2.17s/ examples]

Unsloth: Tokenizing ["text"] (num_proc=12):  17%|█▋        | 2/12 [00:03<00:16,  1.69s/ examples]

Unsloth: Tokenizing ["text"] (num_proc=12):  25%|██▌       | 3/12 [00:04<00:13,  1.52s/ examples]

Unsloth: Tokenizing ["text"] (num_proc=12):  33%|███▎      | 4/12 [00:06<00:11,  1.46s/ examples]

Unsloth: Tokenizing ["text"] (num_proc=12):  42%|████▏     | 5/12 [00:07<00:09,  1.40s/ examples]

Unsloth: Tokenizing ["text"] (num_proc=12):  50%|█████     | 6/12 [00:08<00:08,  1.38s/ examples]

Unsloth: Tokenizing ["text"] (num_proc=12):  58%|█████▊    | 7/12 [00:10<00:06,  1.37s/ examples]

Unsloth: Tokenizing ["text"] (num_proc=12):  67%|██████▋   | 8/12 [00:11<00:05,  1.37s/ examples]

Unsloth: Tokenizing ["text"] (num_proc=12):  75%|███████▌  | 9/12 [00:12<00:04,  1.36s/ examples]

Unsloth: Tokenizing ["text"] (num_proc=12):  83%|████████▎ | 10/12 [00:14<00:02,  1.35s/ examples]

Unsloth: Tokenizing ["text"] (num_proc=12):  92%|█████████▏| 11/12 [00:15<00:01,  1.34s/ examples]

Unsloth: Tokenizing ["text"] (num_proc=12): 100%|██████████| 12/12 [00:16<00:00,  1.33s/ examples]
Unsloth: Tokenizing ["text"] (num_proc=12): 100%|██████████| 12/12 [00:17<00:00,  1.43s/ examples]
🦥 Unsloth: Padding-free auto-enabled, enabling faster training.
The tokenizer has new PAD/BOS/EOS tokens that differ from the model config and generation config. The model config and generation config were aligned accordingly, being updated with the tokenizer's values. Updated tokens: {'bos_token_id': None}.
==((====))==  Unsloth - 2x faster free finetuning | Num GPUs used = 1
   \\   /|    Num examples = 12 | Num Epochs = 3 | Total steps = 18
O^O/ \_/ \    Batch size per device = 1 | Gradient accumulation steps = 2
\        /    Data Parallel GPUs = 1 | Total batch size (1 x 2 x 1) = 2
 "-____-"     Trainable parameters = 40,370,176 of 7,655,986,688 (0.53% trained)


  0%|          | 0/18 [00:00<?, ?it/s]`use_return_dict` is deprecated! Use `return_dict` instead!


  6%|▌         | 1/18 [00:03<01:07,  3.98s/it]


{'loss': '1.994', 'grad_norm': '0.5622', 'learning_rate': '5e-05', 'epoch': '0.3333'}


 11%|█         | 2/18 [00:04<01:03,  3.98s/it]

 17%|█▋        | 3/18 [00:05<00:26,  1.75s/it]

 22%|██▏       | 4/18 [00:07<00:21,  1.54s/it]


{'loss': '1.965', 'grad_norm': '0.4948', 'learning_rate': '4.831e-05', 'epoch': '0.6667'}


 22%|██▏       | 4/18 [00:07<00:21,  1.54s/it]

 28%|██▊       | 5/18 [00:08<00:17,  1.36s/it]


{'loss': '1.788', 'grad_norm': '0.469', 'learning_rate': '4.348e-05', 'epoch': '1'}


 33%|███▎      | 6/18 [00:09<00:16,  1.36s/it]

 39%|███▉      | 7/18 [00:10<00:13,  1.20s/it]


{'loss': '1.795', 'grad_norm': '0.4098', 'learning_rate': '3.614e-05', 'epoch': '1.333'}


 44%|████▍     | 8/18 [00:11<00:11,  1.20s/it]

 50%|█████     | 9/18 [00:11<00:09,  1.07s/it]

 56%|█████▌    | 10/18 [00:12<00:08,  1.07s/it]


{'loss': '1.72', 'grad_norm': '0.37', 'learning_rate': '2.731e-05', 'epoch': '1.667'}


 56%|█████▌    | 10/18 [00:12<00:08,  1.07s/it]

 67%|██████▋   | 12/18 [00:14<00:06,  1.04s/it]


{'loss': '1.659', 'grad_norm': '0.3249', 'learning_rate': '1.816e-05', 'epoch': '2'}


 67%|██████▋   | 12/18 [00:14<00:06,  1.04s/it]

 72%|███████▏  | 13/18 [00:16<00:05,  1.04s/it]

 78%|███████▊  | 14/18 [00:17<00:04,  1.03s/it]


{'loss': '1.68', 'grad_norm': '0.3444', 'learning_rate': '9.934e-06', 'epoch': '2.333'}


 78%|███████▊  | 14/18 [00:17<00:04,  1.03s/it]

 89%|████████▉ | 16/18 [00:18<00:02,  1.02s/it]


{'loss': '1.66', 'grad_norm': '0.2995', 'learning_rate': '3.745e-06', 'epoch': '2.667'}


 89%|████████▉ | 16/18 [00:18<00:02,  1.02s/it]

100%|██████████| 18/18 [00:20<00:00,  1.02it/s]


{'loss': '1.625', 'grad_norm': '0.3541', 'learning_rate': '4.257e-07', 'epoch': '3'}


100%|██████████| 18/18 [00:20<00:00,  1.02it/s]Unsloth: Restored added_tokens_decoder metadata in ./securereview-sft/checkpoint-18/tokenizer_config.json.



{'train_runtime': '21.32', 'train_samples_per_second': '1.688', 'train_steps_per_second': '0.844', 'train_loss': '1.765', 'epoch': '3'}


100%|██████████| 18/18 [00:21<00:00,  1.02it/s]
100%|██████████| 18/18 [00:21<00:00,  1.18s/it]

[6/6] Post-SFT evaluation...
Both `max_new_tokens` (=600) and `max_length`(=32768) seem to have been set. `max_new_tokens` will take precedence. Please refer to the documentation for more information. (https://huggingface.co/docs/transformers/main/en/main_classes/text_generation)
/root/.pyenv/versions/3.13.13/lib/python3.13/site-packages/transformers/modeling_attn_mask_utils.py:71: FutureWarning: The attention mask API under `transformers.modeling_attn_mask_utils` (`AttentionMaskConverter`) is deprecated and will be removed in Transformers v5.10. Please use the new API in `transformers.masking_utils`.
  warnings.warn(DEPRECATION_MESSAGE, FutureWarning)
/root/.pyenv/versions/3.13.13/lib/python3.13/site-packages/transformers/modeling_attn_mask_utils.py:281: FutureWarning: The attention mask API under `transformers.modeling_attn_mask_utils` (`AttentionMaskConverter`) is deprecated and will be removed in Transformers v5.10. Please use the new API in `transformers.masking_utils`.
  warnings.warn(DEPRECATION_MESSAGE, FutureWarning)
  [after] migration_002: 0.300
Both `max_new_tokens` (=600) and `max_length`(=32768) seem to have been set. `max_new_tokens` will take precedence. Please refer to the documentation for more information. (https://huggingface.co/docs/transformers/main/en/main_classes/text_generation)
  [after] migration_006: 0.640
Both `max_new_tokens` (=600) and `max_length`(=32768) seem to have been set. `max_new_tokens` will take precedence. Please refer to the documentation for more information. (https://huggingface.co/docs/transformers/main/en/main_classes/text_generation)
  [after] migration_007: 0.610
Both `max_new_tokens` (=600) and `max_length`(=32768) seem to have been set. `max_new_tokens` will take precedence. Please refer to the documentation for more information. (https://huggingface.co/docs/transformers/main/en/main_classes/text_generation)
  [after] migration_009: 0.200
Both `max_new_tokens` (=600) and `max_length`(=32768) seem to have been set. `max_new_tokens` will take precedence. Please refer to the documentation for more information. (https://huggingface.co/docs/transformers/main/en/main_classes/text_generation)
  [after] migration_012: 0.470
Both `max_new_tokens` (=600) and `max_length`(=32768) seem to have been set. `max_new_tokens` will take precedence. Please refer to the documentation for more information. (https://huggingface.co/docs/transformers/main/en/main_classes/text_generation)
  [after] migration_017: 0.520
Both `max_new_tokens` (=600) and `max_length`(=32768) seem to have been set. `max_new_tokens` will take precedence. Please refer to the documentation for more information. (https://huggingface.co/docs/transformers/main/en/main_classes/text_generation)
  [after] migration_018: 0.520
Both `max_new_tokens` (=600) and `max_length`(=32768) seem to have been set. `max_new_tokens` will take precedence. Please refer to the documentation for more information. (https://huggingface.co/docs/transformers/main/en/main_classes/text_generation)
  [after] migration_022: 0.440
Both `max_new_tokens` (=600) and `max_length`(=32768) seem to have been set. `max_new_tokens` will take precedence. Please refer to the documentation for more information. (https://huggingface.co/docs/transformers/main/en/main_classes/text_generation)
  [after] migration_023: 0.440
Both `max_new_tokens` (=600) and `max_length`(=32768) seem to have been set. `max_new_tokens` will take precedence. Please refer to the documentation for more information. (https://huggingface.co/docs/transformers/main/en/main_classes/text_generation)
  [after] migration_024: 0.330
Both `max_new_tokens` (=600) and `max_length`(=32768) seem to have been set. `max_new_tokens` will take precedence. Please refer to the documentation for more information. (https://huggingface.co/docs/transformers/main/en/main_classes/text_generation)
  [after] migration_025: 0.640
Both `max_new_tokens` (=600) and `max_length`(=32768) seem to have been set. `max_new_tokens` will take precedence. Please refer to the documentation for more information. (https://huggingface.co/docs/transformers/main/en/main_classes/text_generation)
  [after] migration_028: 0.470
  Trained mean: 0.465

=== Improvement Summary ===
  migration_002: 0.300 → 0.300  — +0.000
  migration_006: 0.520 → 0.640  ▲ +0.120
  migration_007: 0.060 → 0.610  ▲ +0.550
  migration_009: 0.260 → 0.200  ▼ -0.060
  migration_012: 0.060 → 0.470  ▲ +0.410
  migration_017: 0.060 → 0.520  ▲ +0.460
  migration_018: 0.290 → 0.520  ▲ +0.230
  migration_022: 0.060 → 0.440  ▲ +0.380
  migration_023: 0.060 → 0.440  ▲ +0.380
  migration_024: 0.280 → 0.330  ▲ +0.050
  migration_025: 0.060 → 0.640  ▲ +0.580
  migration_028: 0.030 → 0.470  ▲ +0.440
  Saved ./plots/reward_curve.png
  Saved ./plots/before_after.png

============================================================
  DONE — Mean 0.170 → 0.465
============================================================