add phi4 models

runs/slm_decoder_models/phi-4/phi-4-phi4_classification_lora-C1/.hydra/config.yaml

@@ -0,0 +1,37 @@
+cache_dir: /tmp/
+dataset:
+  name: kamel-usp/aes_enem_dataset
+  split: JBCS2025
+training_params:
+  seed: 42
+  num_train_epochs: 20
+  logging_steps: 100
+  metric_for_best_model: QWK
+  bf16: true
+post_training_results:
+  model_path: /workspace/jbcs2025/outputs/2025-03-24/20-42-59
+experiments:
+  model:
+    name: microsoft/phi-4
+    type: phi4_classification_lora
+    num_labels: 6
+    output_dir: ./results/phi4-balanced/C1
+    logging_dir: ./logs/phi4-balanced/C1
+    best_model_dir: ./results/phi4-balanced/C1/best_model
+    lora_r: 8
+    lora_dropout: 0.05
+    lora_alpha: 16
+    lora_target_modules: all-linear
+    checkpoint_path: kamel-usp/jbcs2025_phi4-balanced-C1
+  tokenizer:
+    name: microsoft/phi-4
+  dataset:
+    grade_index: 0
+  training_params:
+    weight_decay: 0.01
+    warmup_ratio: 0.1
+    learning_rate: 5.0e-05
+    train_batch_size: 4
+    eval_batch_size: 4
+    gradient_accumulation_steps: 4
+    gradient_checkpointing: false

runs/slm_decoder_models/phi-4/phi-4-phi4_classification_lora-C1/.hydra/hydra.yaml

@@ -0,0 +1,155 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: run_inference_experiment
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: config
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.1'
+    cwd: /workspace/jbcs2025
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /workspace/jbcs2025/configs
+      schema: file
+      provider: main
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /workspace/jbcs2025/outputs/2025-05-26/15-16-56
+    choices:
+      experiments: slm_decoder_models/C1
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false

runs/slm_decoder_models/phi-4/phi-4-phi4_classification_lora-C1/.hydra/overrides.yaml

@@ -0,0 +1 @@
+[]

runs/slm_decoder_models/phi-4/phi-4-phi4_classification_lora-C1/evaluation_results.csv

@@ -0,0 +1,2 @@
+accuracy,RMSE,QWK,HDIV,Macro_F1,Micro_F1,Weighted_F1,TP_0,TN_0,FP_0,FN_0,TP_1,TN_1,FP_1,FN_1,TP_2,TN_2,FP_2,FN_2,TP_3,TN_3,FP_3,FN_3,TP_4,TN_4,FP_4,FN_4,TP_5,TN_5,FP_5,FN_5,timestamp,id
+0.6014492753623188,27.662562992324986,0.6652455160246986,0.007246376811594235,0.4580067731011127,0.6014492753623188,0.6013651374602975,0,137,0,1,0,138,0,0,8,114,14,2,31,63,9,35,37,64,23,14,7,119,9,3,2025-05-26 15:16:56,phi-4-phi4_classification_lora-C1

runs/slm_decoder_models/phi-4/phi-4-phi4_classification_lora-C1/run_inference_experiment.log

@@ -0,0 +1,124 @@
+[2025-05-26 15:16:56,821][__main__][INFO] - Starting inference experiment
+[2025-05-26 15:16:56,823][__main__][INFO] - cache_dir: /tmp/
+dataset:
+  name: kamel-usp/aes_enem_dataset
+  split: JBCS2025
+training_params:
+  seed: 42
+  num_train_epochs: 20
+  logging_steps: 100
+  metric_for_best_model: QWK
+  bf16: true
+post_training_results:
+  model_path: /workspace/jbcs2025/outputs/2025-03-24/20-42-59
+experiments:
+  model:
+    name: microsoft/phi-4
+    type: phi4_classification_lora
+    num_labels: 6
+    output_dir: ./results/phi4-balanced/C1
+    logging_dir: ./logs/phi4-balanced/C1
+    best_model_dir: ./results/phi4-balanced/C1/best_model
+    lora_r: 8
+    lora_dropout: 0.05
+    lora_alpha: 16
+    lora_target_modules: all-linear
+    checkpoint_path: kamel-usp/jbcs2025_phi4-balanced-C1
+  tokenizer:
+    name: microsoft/phi-4
+  dataset:
+    grade_index: 0
+  training_params:
+    weight_decay: 0.01
+    warmup_ratio: 0.1
+    learning_rate: 5.0e-05
+    train_batch_size: 4
+    eval_batch_size: 4
+    gradient_accumulation_steps: 4
+    gradient_checkpointing: false
+
+[2025-05-26 15:16:56,826][__main__][INFO] - Running inference with fine-tuned HF model
+[2025-05-26 15:17:10,431][transformers.tokenization_utils_base][INFO] - loading file vocab.json from cache at /tmp/models--microsoft--phi-4/snapshots/187ef0342fff0eb3333be9f00389385e95ef0b61/vocab.json
+[2025-05-26 15:17:10,432][transformers.tokenization_utils_base][INFO] - loading file merges.txt from cache at /tmp/models--microsoft--phi-4/snapshots/187ef0342fff0eb3333be9f00389385e95ef0b61/merges.txt
+[2025-05-26 15:17:10,432][transformers.tokenization_utils_base][INFO] - loading file tokenizer.json from cache at /tmp/models--microsoft--phi-4/snapshots/187ef0342fff0eb3333be9f00389385e95ef0b61/tokenizer.json
+[2025-05-26 15:17:10,432][transformers.tokenization_utils_base][INFO] - loading file added_tokens.json from cache at /tmp/models--microsoft--phi-4/snapshots/187ef0342fff0eb3333be9f00389385e95ef0b61/added_tokens.json
+[2025-05-26 15:17:10,432][transformers.tokenization_utils_base][INFO] - loading file special_tokens_map.json from cache at /tmp/models--microsoft--phi-4/snapshots/187ef0342fff0eb3333be9f00389385e95ef0b61/special_tokens_map.json
+[2025-05-26 15:17:10,433][transformers.tokenization_utils_base][INFO] - loading file tokenizer_config.json from cache at /tmp/models--microsoft--phi-4/snapshots/187ef0342fff0eb3333be9f00389385e95ef0b61/tokenizer_config.json
+[2025-05-26 15:17:10,433][transformers.tokenization_utils_base][INFO] - loading file chat_template.jinja from cache at None
+[2025-05-26 15:17:10,825][__main__][INFO] - Tokenizer function parameters- Padding:longest; Truncation: False
+[2025-05-26 15:17:12,494][__main__][INFO] - Loading model from: microsoft/phi-4
+[2025-05-26 15:17:13,247][transformers.configuration_utils][INFO] - loading configuration file config.json from cache at /tmp/models--microsoft--phi-4/snapshots/187ef0342fff0eb3333be9f00389385e95ef0b61/config.json
+[2025-05-26 15:17:13,251][transformers.configuration_utils][INFO] - Model config Phi3Config {
+  "architectures": [
+    "Phi3ForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "bos_token_id": 100257,
+  "embd_pdrop": 0.0,
+  "eos_token_id": 100265,
+  "hidden_act": "silu",
+  "hidden_size": 5120,
+  "id2label": {
+    "0": 0,
+    "1": 40,
+    "2": 80,
+    "3": 120,
+    "4": 160,
+    "5": 200
+  },
+  "initializer_range": 0.02,
+  "intermediate_size": 17920,
+  "label2id": {
+    "0": 0,
+    "40": 1,
+    "80": 2,
+    "120": 3,
+    "160": 4,
+    "200": 5
+  },
+  "max_position_embeddings": 16384,
+  "model_type": "phi3",
+  "num_attention_heads": 40,
+  "num_hidden_layers": 40,
+  "num_key_value_heads": 10,
+  "original_max_position_embeddings": 16384,
+  "pad_token_id": 100349,
+  "partial_rotary_factor": 1.0,
+  "resid_pdrop": 0.0,
+  "rms_norm_eps": 1e-05,
+  "rope_scaling": null,
+  "rope_theta": 250000,
+  "sliding_window": null,
+  "tie_word_embeddings": false,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.52.3",
+  "use_cache": true,
+  "vocab_size": 100352
+}
+
+[2025-05-26 15:17:14,641][transformers.modeling_utils][INFO] - loading weights file model.safetensors from cache at /tmp/models--microsoft--phi-4/snapshots/187ef0342fff0eb3333be9f00389385e95ef0b61/model.safetensors.index.json
+[2025-05-26 15:22:17,545][transformers.modeling_utils][INFO] - Will use torch_dtype=torch.bfloat16 as defined in model's config object
+[2025-05-26 15:22:17,546][transformers.modeling_utils][INFO] - Instantiating Phi3ForSequenceClassification model under default dtype torch.bfloat16.
+[2025-05-26 15:22:26,719][transformers.modeling_utils][INFO] - Some weights of the model checkpoint at microsoft/phi-4 were not used when initializing Phi3ForSequenceClassification: ['lm_head.weight']
+- This IS expected if you are initializing Phi3ForSequenceClassification from the checkpoint of a model trained on another task or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPreTraining model).
+- This IS NOT expected if you are initializing Phi3ForSequenceClassification from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).
+[2025-05-26 15:22:26,719][transformers.modeling_utils][WARNING] - Some weights of Phi3ForSequenceClassification were not initialized from the model checkpoint at microsoft/phi-4 and are newly initialized: ['score.weight']
+You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
+[2025-05-26 15:22:34,510][__main__][INFO] - Loaded pre-trained PEFT model from kamel-usp/jbcs2025_phi4-balanced-C1
+[2025-05-26 15:22:34,515][__main__][INFO] - None
+[2025-05-26 15:22:34,519][transformers.training_args][INFO] - PyTorch: setting up devices
+[2025-05-26 15:22:34,579][transformers.training_args][INFO] - The default value for the training argument `--report_to` will change in v5 (from all installed integrations to none). In v5, you will need to use `--report_to all` to get the same behavior as now. You should start updating your code and make this info disappear :-).
+[2025-05-26 15:22:34,599][transformers.trainer][INFO] - You have loaded a model on multiple GPUs. `is_model_parallel` attribute will be force-set to `True` to avoid any unexpected behavior such as device placement mismatching.
+[2025-05-26 15:22:34,748][transformers.trainer][INFO] - Using auto half precision backend
+[2025-05-26 15:22:34,749][transformers.trainer][WARNING] - No label_names provided for model class `PeftModelForSequenceClassification`. Since `PeftModel` hides base models input arguments, if label_names is not given, label_names can't be set automatically within `Trainer`. Note that empty label_names list will be used instead.
+[2025-05-26 15:22:34,749][__main__][INFO] - Running inference on test dataset
+[2025-05-26 15:22:34,751][transformers.trainer][INFO] - The following columns in the test set don't have a corresponding argument in `PeftModelForSequenceClassification.forward` and have been ignored: essay_year, essay_text, grades, prompt, reference, supporting_text, id, id_prompt. If essay_year, essay_text, grades, prompt, reference, supporting_text, id, id_prompt are not expected by `PeftModelForSequenceClassification.forward`,  you can safely ignore this message.
+[2025-05-26 15:22:34,783][transformers.trainer][INFO] - 
+***** Running Prediction *****
+[2025-05-26 15:22:34,783][transformers.trainer][INFO] -   Num examples = 138
+[2025-05-26 15:22:34,783][transformers.trainer][INFO] -   Batch size = 4
+[2025-05-26 15:24:20,731][transformers][INFO] - {'accuracy': 0.6014492753623188, 'RMSE': 27.662562992324986, 'QWK': 0.6652455160246986, 'HDIV': 0.007246376811594235, 'Macro_F1': 0.4580067731011127, 'Micro_F1': 0.6014492753623188, 'Weighted_F1': 0.6013651374602975, 'TP_0': np.int64(0), 'TN_0': np.int64(137), 'FP_0': np.int64(0), 'FN_0': np.int64(1), 'TP_1': np.int64(0), 'TN_1': np.int64(138), 'FP_1': np.int64(0), 'FN_1': np.int64(0), 'TP_2': np.int64(8), 'TN_2': np.int64(114), 'FP_2': np.int64(14), 'FN_2': np.int64(2), 'TP_3': np.int64(31), 'TN_3': np.int64(63), 'FP_3': np.int64(9), 'FN_3': np.int64(35), 'TP_4': np.int64(37), 'TN_4': np.int64(64), 'FP_4': np.int64(23), 'FN_4': np.int64(14), 'TP_5': np.int64(7), 'TN_5': np.int64(119), 'FP_5': np.int64(9), 'FN_5': np.int64(3)}
+[2025-05-26 15:24:20,753][__main__][INFO] - Inference results saved to phi-4-phi4_classification_lora-C1_inference_results.jsonl
+[2025-05-26 15:24:20,754][__main__][INFO] - Inference results: {'accuracy': 0.6014492753623188, 'RMSE': 27.662562992324986, 'QWK': 0.6652455160246986, 'HDIV': 0.007246376811594235, 'Macro_F1': 0.4580067731011127, 'Micro_F1': 0.6014492753623188, 'Weighted_F1': 0.6013651374602975, 'TP_0': np.int64(0), 'TN_0': np.int64(137), 'FP_0': np.int64(0), 'FN_0': np.int64(1), 'TP_1': np.int64(0), 'TN_1': np.int64(138), 'FP_1': np.int64(0), 'FN_1': np.int64(0), 'TP_2': np.int64(8), 'TN_2': np.int64(114), 'FP_2': np.int64(14), 'FN_2': np.int64(2), 'TP_3': np.int64(31), 'TN_3': np.int64(63), 'FP_3': np.int64(9), 'FN_3': np.int64(35), 'TP_4': np.int64(37), 'TN_4': np.int64(64), 'FP_4': np.int64(23), 'FN_4': np.int64(14), 'TP_5': np.int64(7), 'TN_5': np.int64(119), 'FP_5': np.int64(9), 'FN_5': np.int64(3)}
+[2025-05-26 15:24:20,754][__main__][INFO] - Inference experiment completed

runs/slm_decoder_models/phi-4/phi-4-phi4_classification_lora-C2/.hydra/config.yaml

@@ -0,0 +1,37 @@
+cache_dir: /tmp/
+dataset:
+  name: kamel-usp/aes_enem_dataset
+  split: JBCS2025
+training_params:
+  seed: 42
+  num_train_epochs: 20
+  logging_steps: 100
+  metric_for_best_model: QWK
+  bf16: true
+post_training_results:
+  model_path: /workspace/jbcs2025/outputs/2025-03-24/20-42-59
+experiments:
+  model:
+    name: microsoft/phi-4
+    type: phi4_classification_lora
+    num_labels: 6
+    output_dir: ./results/phi4-balanced/C2
+    logging_dir: ./logs/phi4-balanced/C2
+    best_model_dir: ./results/phi4-balanced/C2/best_model
+    lora_r: 8
+    lora_dropout: 0.05
+    lora_alpha: 16
+    lora_target_modules: all-linear
+    checkpoint_path: kamel-usp/jbcs2025_phi4-balanced-C2
+  tokenizer:
+    name: microsoft/phi-4
+  dataset:
+    grade_index: 1
+  training_params:
+    weight_decay: 0.01
+    warmup_ratio: 0.1
+    learning_rate: 5.0e-05
+    train_batch_size: 1
+    eval_batch_size: 4
+    gradient_accumulation_steps: 16
+    gradient_checkpointing: false

runs/slm_decoder_models/phi-4/phi-4-phi4_classification_lora-C2/.hydra/hydra.yaml

@@ -0,0 +1,155 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: run_inference_experiment
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: config
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.1'
+    cwd: /workspace/jbcs2025
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /workspace/jbcs2025/configs
+      schema: file
+      provider: main
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /workspace/jbcs2025/outputs/2025-05-26/15-50-03
+    choices:
+      experiments: slm_decoder_models/C2
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false

runs/slm_decoder_models/phi-4/phi-4-phi4_classification_lora-C2/.hydra/overrides.yaml

@@ -0,0 +1 @@
+[]

runs/slm_decoder_models/phi-4/phi-4-phi4_classification_lora-C2/evaluation_results.csv

@@ -0,0 +1,2 @@
+accuracy,RMSE,QWK,HDIV,Macro_F1,Micro_F1,Weighted_F1,TP_0,TN_0,FP_0,FN_0,TP_1,TN_1,FP_1,FN_1,TP_2,TN_2,FP_2,FN_2,TP_3,TN_3,FP_3,FN_3,TP_4,TN_4,FP_4,FN_4,TP_5,TN_5,FP_5,FN_5,timestamp,id
+0.47101449275362317,61.2904702146299,0.39747736093143593,0.08695652173913049,0.28940811083093393,0.47101449275362317,0.44952547422123834,0,137,0,1,21,78,25,14,0,133,0,5,31,60,27,20,6,106,6,20,7,103,15,13,2025-05-26 15:50:03,phi-4-phi4_classification_lora-C2

runs/slm_decoder_models/phi-4/phi-4-phi4_classification_lora-C2/run_inference_experiment.log

@@ -0,0 +1,124 @@
+[2025-05-26 15:50:03,858][__main__][INFO] - Starting inference experiment
+[2025-05-26 15:50:03,860][__main__][INFO] - cache_dir: /tmp/
+dataset:
+  name: kamel-usp/aes_enem_dataset
+  split: JBCS2025
+training_params:
+  seed: 42
+  num_train_epochs: 20
+  logging_steps: 100
+  metric_for_best_model: QWK
+  bf16: true
+post_training_results:
+  model_path: /workspace/jbcs2025/outputs/2025-03-24/20-42-59
+experiments:
+  model:
+    name: microsoft/phi-4
+    type: phi4_classification_lora
+    num_labels: 6
+    output_dir: ./results/phi4-balanced/C2
+    logging_dir: ./logs/phi4-balanced/C2
+    best_model_dir: ./results/phi4-balanced/C2/best_model
+    lora_r: 8
+    lora_dropout: 0.05
+    lora_alpha: 16
+    lora_target_modules: all-linear
+    checkpoint_path: kamel-usp/jbcs2025_phi4-balanced-C2
+  tokenizer:
+    name: microsoft/phi-4
+  dataset:
+    grade_index: 1
+  training_params:
+    weight_decay: 0.01
+    warmup_ratio: 0.1
+    learning_rate: 5.0e-05
+    train_batch_size: 1
+    eval_batch_size: 4
+    gradient_accumulation_steps: 16
+    gradient_checkpointing: false
+
+[2025-05-26 15:50:03,863][__main__][INFO] - Running inference with fine-tuned HF model
+[2025-05-26 15:50:11,781][transformers.tokenization_utils_base][INFO] - loading file vocab.json from cache at /tmp/models--microsoft--phi-4/snapshots/187ef0342fff0eb3333be9f00389385e95ef0b61/vocab.json
+[2025-05-26 15:50:11,782][transformers.tokenization_utils_base][INFO] - loading file merges.txt from cache at /tmp/models--microsoft--phi-4/snapshots/187ef0342fff0eb3333be9f00389385e95ef0b61/merges.txt
+[2025-05-26 15:50:11,782][transformers.tokenization_utils_base][INFO] - loading file tokenizer.json from cache at /tmp/models--microsoft--phi-4/snapshots/187ef0342fff0eb3333be9f00389385e95ef0b61/tokenizer.json
+[2025-05-26 15:50:11,782][transformers.tokenization_utils_base][INFO] - loading file added_tokens.json from cache at /tmp/models--microsoft--phi-4/snapshots/187ef0342fff0eb3333be9f00389385e95ef0b61/added_tokens.json
+[2025-05-26 15:50:11,782][transformers.tokenization_utils_base][INFO] - loading file special_tokens_map.json from cache at /tmp/models--microsoft--phi-4/snapshots/187ef0342fff0eb3333be9f00389385e95ef0b61/special_tokens_map.json
+[2025-05-26 15:50:11,782][transformers.tokenization_utils_base][INFO] - loading file tokenizer_config.json from cache at /tmp/models--microsoft--phi-4/snapshots/187ef0342fff0eb3333be9f00389385e95ef0b61/tokenizer_config.json
+[2025-05-26 15:50:11,783][transformers.tokenization_utils_base][INFO] - loading file chat_template.jinja from cache at None
+[2025-05-26 15:50:12,042][__main__][INFO] - Tokenizer function parameters- Padding:longest; Truncation: False
+[2025-05-26 15:50:12,617][__main__][INFO] - Loading model from: microsoft/phi-4
+[2025-05-26 15:50:12,863][transformers.configuration_utils][INFO] - loading configuration file config.json from cache at /tmp/models--microsoft--phi-4/snapshots/187ef0342fff0eb3333be9f00389385e95ef0b61/config.json
+[2025-05-26 15:50:12,866][transformers.configuration_utils][INFO] - Model config Phi3Config {
+  "architectures": [
+    "Phi3ForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "bos_token_id": 100257,
+  "embd_pdrop": 0.0,
+  "eos_token_id": 100265,
+  "hidden_act": "silu",
+  "hidden_size": 5120,
+  "id2label": {
+    "0": 0,
+    "1": 40,
+    "2": 80,
+    "3": 120,
+    "4": 160,
+    "5": 200
+  },
+  "initializer_range": 0.02,
+  "intermediate_size": 17920,
+  "label2id": {
+    "0": 0,
+    "40": 1,
+    "80": 2,
+    "120": 3,
+    "160": 4,
+    "200": 5
+  },
+  "max_position_embeddings": 16384,
+  "model_type": "phi3",
+  "num_attention_heads": 40,
+  "num_hidden_layers": 40,
+  "num_key_value_heads": 10,
+  "original_max_position_embeddings": 16384,
+  "pad_token_id": 100349,
+  "partial_rotary_factor": 1.0,
+  "resid_pdrop": 0.0,
+  "rms_norm_eps": 1e-05,
+  "rope_scaling": null,
+  "rope_theta": 250000,
+  "sliding_window": null,
+  "tie_word_embeddings": false,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.52.3",
+  "use_cache": true,
+  "vocab_size": 100352
+}
+
+[2025-05-26 15:50:13,103][transformers.modeling_utils][INFO] - loading weights file model.safetensors from cache at /tmp/models--microsoft--phi-4/snapshots/187ef0342fff0eb3333be9f00389385e95ef0b61/model.safetensors.index.json
+[2025-05-26 15:50:13,104][transformers.modeling_utils][INFO] - Will use torch_dtype=torch.bfloat16 as defined in model's config object
+[2025-05-26 15:50:13,104][transformers.modeling_utils][INFO] - Instantiating Phi3ForSequenceClassification model under default dtype torch.bfloat16.
+[2025-05-26 15:50:21,981][transformers.modeling_utils][INFO] - Some weights of the model checkpoint at microsoft/phi-4 were not used when initializing Phi3ForSequenceClassification: ['lm_head.weight']
+- This IS expected if you are initializing Phi3ForSequenceClassification from the checkpoint of a model trained on another task or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPreTraining model).
+- This IS NOT expected if you are initializing Phi3ForSequenceClassification from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).
+[2025-05-26 15:50:21,981][transformers.modeling_utils][WARNING] - Some weights of Phi3ForSequenceClassification were not initialized from the model checkpoint at microsoft/phi-4 and are newly initialized: ['score.weight']
+You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
+[2025-05-26 15:50:24,590][__main__][INFO] - Loaded pre-trained PEFT model from kamel-usp/jbcs2025_phi4-balanced-C2
+[2025-05-26 15:50:24,595][__main__][INFO] - None
+[2025-05-26 15:50:24,598][transformers.training_args][INFO] - PyTorch: setting up devices
+[2025-05-26 15:50:24,649][transformers.training_args][INFO] - The default value for the training argument `--report_to` will change in v5 (from all installed integrations to none). In v5, you will need to use `--report_to all` to get the same behavior as now. You should start updating your code and make this info disappear :-).
+[2025-05-26 15:50:24,665][transformers.trainer][INFO] - You have loaded a model on multiple GPUs. `is_model_parallel` attribute will be force-set to `True` to avoid any unexpected behavior such as device placement mismatching.
+[2025-05-26 15:50:24,700][transformers.trainer][INFO] - Using auto half precision backend
+[2025-05-26 15:50:24,701][transformers.trainer][WARNING] - No label_names provided for model class `PeftModelForSequenceClassification`. Since `PeftModel` hides base models input arguments, if label_names is not given, label_names can't be set automatically within `Trainer`. Note that empty label_names list will be used instead.
+[2025-05-26 15:50:24,701][__main__][INFO] - Running inference on test dataset
+[2025-05-26 15:50:24,702][transformers.trainer][INFO] - The following columns in the test set don't have a corresponding argument in `PeftModelForSequenceClassification.forward` and have been ignored: reference, essay_year, id_prompt, supporting_text, prompt, id, essay_text, grades. If reference, essay_year, id_prompt, supporting_text, prompt, id, essay_text, grades are not expected by `PeftModelForSequenceClassification.forward`,  you can safely ignore this message.
+[2025-05-26 15:50:24,730][transformers.trainer][INFO] - 
+***** Running Prediction *****
+[2025-05-26 15:50:24,730][transformers.trainer][INFO] -   Num examples = 138
+[2025-05-26 15:50:24,730][transformers.trainer][INFO] -   Batch size = 4
+[2025-05-26 15:52:59,219][transformers][INFO] - {'accuracy': 0.47101449275362317, 'RMSE': 61.2904702146299, 'QWK': 0.39747736093143593, 'HDIV': 0.08695652173913049, 'Macro_F1': 0.28940811083093393, 'Micro_F1': 0.47101449275362317, 'Weighted_F1': 0.44952547422123834, 'TP_0': np.int64(0), 'TN_0': np.int64(137), 'FP_0': np.int64(0), 'FN_0': np.int64(1), 'TP_1': np.int64(21), 'TN_1': np.int64(78), 'FP_1': np.int64(25), 'FN_1': np.int64(14), 'TP_2': np.int64(0), 'TN_2': np.int64(133), 'FP_2': np.int64(0), 'FN_2': np.int64(5), 'TP_3': np.int64(31), 'TN_3': np.int64(60), 'FP_3': np.int64(27), 'FN_3': np.int64(20), 'TP_4': np.int64(6), 'TN_4': np.int64(106), 'FP_4': np.int64(6), 'FN_4': np.int64(20), 'TP_5': np.int64(7), 'TN_5': np.int64(103), 'FP_5': np.int64(15), 'FN_5': np.int64(13)}
+[2025-05-26 15:52:59,240][__main__][INFO] - Inference results saved to phi-4-phi4_classification_lora-C2_inference_results.jsonl
+[2025-05-26 15:52:59,241][__main__][INFO] - Inference results: {'accuracy': 0.47101449275362317, 'RMSE': 61.2904702146299, 'QWK': 0.39747736093143593, 'HDIV': 0.08695652173913049, 'Macro_F1': 0.28940811083093393, 'Micro_F1': 0.47101449275362317, 'Weighted_F1': 0.44952547422123834, 'TP_0': np.int64(0), 'TN_0': np.int64(137), 'FP_0': np.int64(0), 'FN_0': np.int64(1), 'TP_1': np.int64(21), 'TN_1': np.int64(78), 'FP_1': np.int64(25), 'FN_1': np.int64(14), 'TP_2': np.int64(0), 'TN_2': np.int64(133), 'FP_2': np.int64(0), 'FN_2': np.int64(5), 'TP_3': np.int64(31), 'TN_3': np.int64(60), 'FP_3': np.int64(27), 'FN_3': np.int64(20), 'TP_4': np.int64(6), 'TN_4': np.int64(106), 'FP_4': np.int64(6), 'FN_4': np.int64(20), 'TP_5': np.int64(7), 'TN_5': np.int64(103), 'FP_5': np.int64(15), 'FN_5': np.int64(13)}
+[2025-05-26 15:52:59,241][__main__][INFO] - Inference experiment completed

runs/slm_decoder_models/phi-4/phi-4-phi4_classification_lora-C3/.hydra/config.yaml

@@ -0,0 +1,37 @@
+cache_dir: /tmp/
+dataset:
+  name: kamel-usp/aes_enem_dataset
+  split: JBCS2025
+training_params:
+  seed: 42
+  num_train_epochs: 20
+  logging_steps: 100
+  metric_for_best_model: QWK
+  bf16: true
+post_training_results:
+  model_path: /workspace/jbcs2025/outputs/2025-03-24/20-42-59
+experiments:
+  model:
+    name: microsoft/phi-4
+    type: phi4_classification_lora
+    num_labels: 6
+    output_dir: ./results/phi4-balanced/C3
+    logging_dir: ./logs/phi4-balanced/C3
+    best_model_dir: ./results/phi4-balanced/C3/best_model
+    lora_r: 8
+    lora_dropout: 0.05
+    lora_alpha: 16
+    lora_target_modules: all-linear
+    checkpoint_path: kamel-usp/jbcs2025_phi4-balanced-C3
+  tokenizer:
+    name: microsoft/phi-4
+  dataset:
+    grade_index: 2
+  training_params:
+    weight_decay: 0.01
+    warmup_ratio: 0.1
+    learning_rate: 5.0e-05
+    train_batch_size: 1
+    eval_batch_size: 2
+    gradient_accumulation_steps: 16
+    gradient_checkpointing: true

runs/slm_decoder_models/phi-4/phi-4-phi4_classification_lora-C3/.hydra/hydra.yaml

@@ -0,0 +1,155 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: run_inference_experiment
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: config
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.1'
+    cwd: /workspace/jbcs2025
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /workspace/jbcs2025/configs
+      schema: file
+      provider: main
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /workspace/jbcs2025/outputs/2025-05-26/15-54-03
+    choices:
+      experiments: slm_decoder_models/C3
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false

runs/slm_decoder_models/phi-4/phi-4-phi4_classification_lora-C3/.hydra/overrides.yaml

@@ -0,0 +1 @@
+[]

runs/slm_decoder_models/phi-4/phi-4-phi4_classification_lora-C3/evaluation_results.csv

@@ -0,0 +1,2 @@
+accuracy,RMSE,QWK,HDIV,Macro_F1,Micro_F1,Weighted_F1,TP_0,TN_0,FP_0,FN_0,TP_1,TN_1,FP_1,FN_1,TP_2,TN_2,FP_2,FN_2,TP_3,TN_3,FP_3,FN_3,TP_4,TN_4,FP_4,FN_4,TP_5,TN_5,FP_5,FN_5,timestamp,id
+0.3333333333333333,49.22589162013322,0.29381519024533553,0.036231884057971064,0.20859754797807895,0.3333333333333333,0.2747500757119841,0,137,0,1,0,109,0,29,14,86,34,4,24,49,44,21,7,93,7,31,1,124,7,6,2025-05-26 15:54:03,phi-4-phi4_classification_lora-C3

runs/slm_decoder_models/phi-4/phi-4-phi4_classification_lora-C3/run_inference_experiment.log

@@ -0,0 +1,124 @@
+[2025-05-26 15:54:03,983][__main__][INFO] - Starting inference experiment
+[2025-05-26 15:54:03,986][__main__][INFO] - cache_dir: /tmp/
+dataset:
+  name: kamel-usp/aes_enem_dataset
+  split: JBCS2025
+training_params:
+  seed: 42
+  num_train_epochs: 20
+  logging_steps: 100
+  metric_for_best_model: QWK
+  bf16: true
+post_training_results:
+  model_path: /workspace/jbcs2025/outputs/2025-03-24/20-42-59
+experiments:
+  model:
+    name: microsoft/phi-4
+    type: phi4_classification_lora
+    num_labels: 6
+    output_dir: ./results/phi4-balanced/C3
+    logging_dir: ./logs/phi4-balanced/C3
+    best_model_dir: ./results/phi4-balanced/C3/best_model
+    lora_r: 8
+    lora_dropout: 0.05
+    lora_alpha: 16
+    lora_target_modules: all-linear
+    checkpoint_path: kamel-usp/jbcs2025_phi4-balanced-C3
+  tokenizer:
+    name: microsoft/phi-4
+  dataset:
+    grade_index: 2
+  training_params:
+    weight_decay: 0.01
+    warmup_ratio: 0.1
+    learning_rate: 5.0e-05
+    train_batch_size: 1
+    eval_batch_size: 2
+    gradient_accumulation_steps: 16
+    gradient_checkpointing: true
+
+[2025-05-26 15:54:03,989][__main__][INFO] - Running inference with fine-tuned HF model
+[2025-05-26 15:54:10,518][transformers.tokenization_utils_base][INFO] - loading file vocab.json from cache at /tmp/models--microsoft--phi-4/snapshots/187ef0342fff0eb3333be9f00389385e95ef0b61/vocab.json
+[2025-05-26 15:54:10,518][transformers.tokenization_utils_base][INFO] - loading file merges.txt from cache at /tmp/models--microsoft--phi-4/snapshots/187ef0342fff0eb3333be9f00389385e95ef0b61/merges.txt
+[2025-05-26 15:54:10,518][transformers.tokenization_utils_base][INFO] - loading file tokenizer.json from cache at /tmp/models--microsoft--phi-4/snapshots/187ef0342fff0eb3333be9f00389385e95ef0b61/tokenizer.json
+[2025-05-26 15:54:10,518][transformers.tokenization_utils_base][INFO] - loading file added_tokens.json from cache at /tmp/models--microsoft--phi-4/snapshots/187ef0342fff0eb3333be9f00389385e95ef0b61/added_tokens.json
+[2025-05-26 15:54:10,518][transformers.tokenization_utils_base][INFO] - loading file special_tokens_map.json from cache at /tmp/models--microsoft--phi-4/snapshots/187ef0342fff0eb3333be9f00389385e95ef0b61/special_tokens_map.json
+[2025-05-26 15:54:10,519][transformers.tokenization_utils_base][INFO] - loading file tokenizer_config.json from cache at /tmp/models--microsoft--phi-4/snapshots/187ef0342fff0eb3333be9f00389385e95ef0b61/tokenizer_config.json
+[2025-05-26 15:54:10,519][transformers.tokenization_utils_base][INFO] - loading file chat_template.jinja from cache at None
+[2025-05-26 15:54:10,885][__main__][INFO] - Tokenizer function parameters- Padding:longest; Truncation: False
+[2025-05-26 15:54:12,790][__main__][INFO] - Loading model from: microsoft/phi-4
+[2025-05-26 15:54:13,140][transformers.configuration_utils][INFO] - loading configuration file config.json from cache at /tmp/models--microsoft--phi-4/snapshots/187ef0342fff0eb3333be9f00389385e95ef0b61/config.json
+[2025-05-26 15:54:13,143][transformers.configuration_utils][INFO] - Model config Phi3Config {
+  "architectures": [
+    "Phi3ForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "bos_token_id": 100257,
+  "embd_pdrop": 0.0,
+  "eos_token_id": 100265,
+  "hidden_act": "silu",
+  "hidden_size": 5120,
+  "id2label": {
+    "0": 0,
+    "1": 40,
+    "2": 80,
+    "3": 120,
+    "4": 160,
+    "5": 200
+  },
+  "initializer_range": 0.02,
+  "intermediate_size": 17920,
+  "label2id": {
+    "0": 0,
+    "40": 1,
+    "80": 2,
+    "120": 3,
+    "160": 4,
+    "200": 5
+  },
+  "max_position_embeddings": 16384,
+  "model_type": "phi3",
+  "num_attention_heads": 40,
+  "num_hidden_layers": 40,
+  "num_key_value_heads": 10,
+  "original_max_position_embeddings": 16384,
+  "pad_token_id": 100349,
+  "partial_rotary_factor": 1.0,
+  "resid_pdrop": 0.0,
+  "rms_norm_eps": 1e-05,
+  "rope_scaling": null,
+  "rope_theta": 250000,
+  "sliding_window": null,
+  "tie_word_embeddings": false,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.52.3",
+  "use_cache": true,
+  "vocab_size": 100352
+}
+
+[2025-05-26 15:54:13,382][transformers.modeling_utils][INFO] - loading weights file model.safetensors from cache at /tmp/models--microsoft--phi-4/snapshots/187ef0342fff0eb3333be9f00389385e95ef0b61/model.safetensors.index.json
+[2025-05-26 15:54:13,382][transformers.modeling_utils][INFO] - Will use torch_dtype=torch.bfloat16 as defined in model's config object
+[2025-05-26 15:54:13,383][transformers.modeling_utils][INFO] - Instantiating Phi3ForSequenceClassification model under default dtype torch.bfloat16.
+[2025-05-26 15:54:22,204][transformers.modeling_utils][INFO] - Some weights of the model checkpoint at microsoft/phi-4 were not used when initializing Phi3ForSequenceClassification: ['lm_head.weight']
+- This IS expected if you are initializing Phi3ForSequenceClassification from the checkpoint of a model trained on another task or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPreTraining model).
+- This IS NOT expected if you are initializing Phi3ForSequenceClassification from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).
+[2025-05-26 15:54:22,205][transformers.modeling_utils][WARNING] - Some weights of Phi3ForSequenceClassification were not initialized from the model checkpoint at microsoft/phi-4 and are newly initialized: ['score.weight']
+You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
+[2025-05-26 15:54:29,389][__main__][INFO] - Loaded pre-trained PEFT model from kamel-usp/jbcs2025_phi4-balanced-C3
+[2025-05-26 15:54:29,395][__main__][INFO] - None
+[2025-05-26 15:54:29,400][transformers.training_args][INFO] - PyTorch: setting up devices
+[2025-05-26 15:54:29,475][transformers.training_args][INFO] - The default value for the training argument `--report_to` will change in v5 (from all installed integrations to none). In v5, you will need to use `--report_to all` to get the same behavior as now. You should start updating your code and make this info disappear :-).
+[2025-05-26 15:54:29,495][transformers.trainer][INFO] - You have loaded a model on multiple GPUs. `is_model_parallel` attribute will be force-set to `True` to avoid any unexpected behavior such as device placement mismatching.
+[2025-05-26 15:54:29,534][transformers.trainer][INFO] - Using auto half precision backend
+[2025-05-26 15:54:29,535][transformers.trainer][WARNING] - No label_names provided for model class `PeftModelForSequenceClassification`. Since `PeftModel` hides base models input arguments, if label_names is not given, label_names can't be set automatically within `Trainer`. Note that empty label_names list will be used instead.
+[2025-05-26 15:54:29,536][__main__][INFO] - Running inference on test dataset
+[2025-05-26 15:54:29,537][transformers.trainer][INFO] - The following columns in the test set don't have a corresponding argument in `PeftModelForSequenceClassification.forward` and have been ignored: grades, reference, essay_year, id, prompt, essay_text, id_prompt, supporting_text. If grades, reference, essay_year, id, prompt, essay_text, id_prompt, supporting_text are not expected by `PeftModelForSequenceClassification.forward`,  you can safely ignore this message.
+[2025-05-26 15:54:29,571][transformers.trainer][INFO] - 
+***** Running Prediction *****
+[2025-05-26 15:54:29,571][transformers.trainer][INFO] -   Num examples = 138
+[2025-05-26 15:54:29,571][transformers.trainer][INFO] -   Batch size = 2
+[2025-05-26 15:56:31,338][transformers][INFO] - {'accuracy': 0.3333333333333333, 'RMSE': 49.22589162013322, 'QWK': 0.29381519024533553, 'HDIV': 0.036231884057971064, 'Macro_F1': 0.20859754797807895, 'Micro_F1': 0.3333333333333333, 'Weighted_F1': 0.2747500757119841, 'TP_0': np.int64(0), 'TN_0': np.int64(137), 'FP_0': np.int64(0), 'FN_0': np.int64(1), 'TP_1': np.int64(0), 'TN_1': np.int64(109), 'FP_1': np.int64(0), 'FN_1': np.int64(29), 'TP_2': np.int64(14), 'TN_2': np.int64(86), 'FP_2': np.int64(34), 'FN_2': np.int64(4), 'TP_3': np.int64(24), 'TN_3': np.int64(49), 'FP_3': np.int64(44), 'FN_3': np.int64(21), 'TP_4': np.int64(7), 'TN_4': np.int64(93), 'FP_4': np.int64(7), 'FN_4': np.int64(31), 'TP_5': np.int64(1), 'TN_5': np.int64(124), 'FP_5': np.int64(7), 'FN_5': np.int64(6)}
+[2025-05-26 15:56:31,364][__main__][INFO] - Inference results saved to phi-4-phi4_classification_lora-C3_inference_results.jsonl
+[2025-05-26 15:56:31,364][__main__][INFO] - Inference results: {'accuracy': 0.3333333333333333, 'RMSE': 49.22589162013322, 'QWK': 0.29381519024533553, 'HDIV': 0.036231884057971064, 'Macro_F1': 0.20859754797807895, 'Micro_F1': 0.3333333333333333, 'Weighted_F1': 0.2747500757119841, 'TP_0': np.int64(0), 'TN_0': np.int64(137), 'FP_0': np.int64(0), 'FN_0': np.int64(1), 'TP_1': np.int64(0), 'TN_1': np.int64(109), 'FP_1': np.int64(0), 'FN_1': np.int64(29), 'TP_2': np.int64(14), 'TN_2': np.int64(86), 'FP_2': np.int64(34), 'FN_2': np.int64(4), 'TP_3': np.int64(24), 'TN_3': np.int64(49), 'FP_3': np.int64(44), 'FN_3': np.int64(21), 'TP_4': np.int64(7), 'TN_4': np.int64(93), 'FP_4': np.int64(7), 'FN_4': np.int64(31), 'TP_5': np.int64(1), 'TN_5': np.int64(124), 'FP_5': np.int64(7), 'FN_5': np.int64(6)}
+[2025-05-26 15:56:31,365][__main__][INFO] - Inference experiment completed

runs/slm_decoder_models/phi-4/phi-4-phi4_classification_lora-C4/.hydra/config.yaml

@@ -0,0 +1,37 @@
+cache_dir: /tmp/
+dataset:
+  name: kamel-usp/aes_enem_dataset
+  split: JBCS2025
+training_params:
+  seed: 42
+  num_train_epochs: 20
+  logging_steps: 100
+  metric_for_best_model: QWK
+  bf16: true
+post_training_results:
+  model_path: /workspace/jbcs2025/outputs/2025-03-24/20-42-59
+experiments:
+  model:
+    name: microsoft/phi-4
+    type: phi4_classification_lora
+    num_labels: 6
+    output_dir: ./results/phi4-balanced/C4
+    logging_dir: ./logs/phi4-balanced/C4
+    best_model_dir: ./results/phi4-balanced/C4/best_model
+    lora_r: 8
+    lora_dropout: 0.05
+    lora_alpha: 16
+    lora_target_modules: all-linear
+    checkpoint_path: kamel-usp/jbcs2025_phi4-balanced-C4
+  tokenizer:
+    name: microsoft/phi-4
+  dataset:
+    grade_index: 3
+  training_params:
+    weight_decay: 0.01
+    warmup_ratio: 0.1
+    learning_rate: 5.0e-05
+    train_batch_size: 2
+    eval_batch_size: 4
+    gradient_accumulation_steps: 8
+    gradient_checkpointing: false

runs/slm_decoder_models/phi-4/phi-4-phi4_classification_lora-C4/.hydra/hydra.yaml

@@ -0,0 +1,155 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: run_inference_experiment
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: config
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.1'
+    cwd: /workspace/jbcs2025
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /workspace/jbcs2025/configs
+      schema: file
+      provider: main
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /workspace/jbcs2025/outputs/2025-05-26/15-57-10
+    choices:
+      experiments: slm_decoder_models/C4
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false

runs/slm_decoder_models/phi-4/phi-4-phi4_classification_lora-C4/.hydra/overrides.yaml

@@ -0,0 +1 @@
+[]

runs/slm_decoder_models/phi-4/phi-4-phi4_classification_lora-C4/evaluation_results.csv

@@ -0,0 +1,2 @@
+accuracy,RMSE,QWK,HDIV,Macro_F1,Micro_F1,Weighted_F1,TP_0,TN_0,FP_0,FN_0,TP_1,TN_1,FP_1,FN_1,TP_2,TN_2,FP_2,FN_2,TP_3,TN_3,FP_3,FN_3,TP_4,TN_4,FP_4,FN_4,TP_5,TN_5,FP_5,FN_5,timestamp,id
+0.7028985507246377,24.55399256179405,0.579465541490858,0.007246376811594235,0.27468966776195697,0.7028985507246377,0.6761155293109196,0,137,0,1,0,135,2,1,1,128,1,8,67,39,23,9,29,79,13,17,0,131,2,5,2025-05-26 15:57:10,phi-4-phi4_classification_lora-C4

runs/slm_decoder_models/phi-4/phi-4-phi4_classification_lora-C4/run_inference_experiment.log

@@ -0,0 +1,124 @@
+[2025-05-26 15:57:10,593][__main__][INFO] - Starting inference experiment
+[2025-05-26 15:57:10,595][__main__][INFO] - cache_dir: /tmp/
+dataset:
+  name: kamel-usp/aes_enem_dataset
+  split: JBCS2025
+training_params:
+  seed: 42
+  num_train_epochs: 20
+  logging_steps: 100
+  metric_for_best_model: QWK
+  bf16: true
+post_training_results:
+  model_path: /workspace/jbcs2025/outputs/2025-03-24/20-42-59
+experiments:
+  model:
+    name: microsoft/phi-4
+    type: phi4_classification_lora
+    num_labels: 6
+    output_dir: ./results/phi4-balanced/C4
+    logging_dir: ./logs/phi4-balanced/C4
+    best_model_dir: ./results/phi4-balanced/C4/best_model
+    lora_r: 8
+    lora_dropout: 0.05
+    lora_alpha: 16
+    lora_target_modules: all-linear
+    checkpoint_path: kamel-usp/jbcs2025_phi4-balanced-C4
+  tokenizer:
+    name: microsoft/phi-4
+  dataset:
+    grade_index: 3
+  training_params:
+    weight_decay: 0.01
+    warmup_ratio: 0.1
+    learning_rate: 5.0e-05
+    train_batch_size: 2
+    eval_batch_size: 4
+    gradient_accumulation_steps: 8
+    gradient_checkpointing: false
+
+[2025-05-26 15:57:10,598][__main__][INFO] - Running inference with fine-tuned HF model
+[2025-05-26 15:57:16,626][transformers.tokenization_utils_base][INFO] - loading file vocab.json from cache at /tmp/models--microsoft--phi-4/snapshots/187ef0342fff0eb3333be9f00389385e95ef0b61/vocab.json
+[2025-05-26 15:57:16,627][transformers.tokenization_utils_base][INFO] - loading file merges.txt from cache at /tmp/models--microsoft--phi-4/snapshots/187ef0342fff0eb3333be9f00389385e95ef0b61/merges.txt
+[2025-05-26 15:57:16,627][transformers.tokenization_utils_base][INFO] - loading file tokenizer.json from cache at /tmp/models--microsoft--phi-4/snapshots/187ef0342fff0eb3333be9f00389385e95ef0b61/tokenizer.json
+[2025-05-26 15:57:16,627][transformers.tokenization_utils_base][INFO] - loading file added_tokens.json from cache at /tmp/models--microsoft--phi-4/snapshots/187ef0342fff0eb3333be9f00389385e95ef0b61/added_tokens.json
+[2025-05-26 15:57:16,627][transformers.tokenization_utils_base][INFO] - loading file special_tokens_map.json from cache at /tmp/models--microsoft--phi-4/snapshots/187ef0342fff0eb3333be9f00389385e95ef0b61/special_tokens_map.json
+[2025-05-26 15:57:16,627][transformers.tokenization_utils_base][INFO] - loading file tokenizer_config.json from cache at /tmp/models--microsoft--phi-4/snapshots/187ef0342fff0eb3333be9f00389385e95ef0b61/tokenizer_config.json
+[2025-05-26 15:57:16,627][transformers.tokenization_utils_base][INFO] - loading file chat_template.jinja from cache at None
+[2025-05-26 15:57:17,010][__main__][INFO] - Tokenizer function parameters- Padding:longest; Truncation: False
+[2025-05-26 15:57:18,888][__main__][INFO] - Loading model from: microsoft/phi-4
+[2025-05-26 15:57:19,197][transformers.configuration_utils][INFO] - loading configuration file config.json from cache at /tmp/models--microsoft--phi-4/snapshots/187ef0342fff0eb3333be9f00389385e95ef0b61/config.json
+[2025-05-26 15:57:19,201][transformers.configuration_utils][INFO] - Model config Phi3Config {
+  "architectures": [
+    "Phi3ForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "bos_token_id": 100257,
+  "embd_pdrop": 0.0,
+  "eos_token_id": 100265,
+  "hidden_act": "silu",
+  "hidden_size": 5120,
+  "id2label": {
+    "0": 0,
+    "1": 40,
+    "2": 80,
+    "3": 120,
+    "4": 160,
+    "5": 200
+  },
+  "initializer_range": 0.02,
+  "intermediate_size": 17920,
+  "label2id": {
+    "0": 0,
+    "40": 1,
+    "80": 2,
+    "120": 3,
+    "160": 4,
+    "200": 5
+  },
+  "max_position_embeddings": 16384,
+  "model_type": "phi3",
+  "num_attention_heads": 40,
+  "num_hidden_layers": 40,
+  "num_key_value_heads": 10,
+  "original_max_position_embeddings": 16384,
+  "pad_token_id": 100349,
+  "partial_rotary_factor": 1.0,
+  "resid_pdrop": 0.0,
+  "rms_norm_eps": 1e-05,
+  "rope_scaling": null,
+  "rope_theta": 250000,
+  "sliding_window": null,
+  "tie_word_embeddings": false,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.52.3",
+  "use_cache": true,
+  "vocab_size": 100352
+}
+
+[2025-05-26 15:57:19,442][transformers.modeling_utils][INFO] - loading weights file model.safetensors from cache at /tmp/models--microsoft--phi-4/snapshots/187ef0342fff0eb3333be9f00389385e95ef0b61/model.safetensors.index.json
+[2025-05-26 15:57:19,443][transformers.modeling_utils][INFO] - Will use torch_dtype=torch.bfloat16 as defined in model's config object
+[2025-05-26 15:57:19,443][transformers.modeling_utils][INFO] - Instantiating Phi3ForSequenceClassification model under default dtype torch.bfloat16.
+[2025-05-26 15:57:28,301][transformers.modeling_utils][INFO] - Some weights of the model checkpoint at microsoft/phi-4 were not used when initializing Phi3ForSequenceClassification: ['lm_head.weight']
+- This IS expected if you are initializing Phi3ForSequenceClassification from the checkpoint of a model trained on another task or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPreTraining model).
+- This IS NOT expected if you are initializing Phi3ForSequenceClassification from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).
+[2025-05-26 15:57:28,302][transformers.modeling_utils][WARNING] - Some weights of Phi3ForSequenceClassification were not initialized from the model checkpoint at microsoft/phi-4 and are newly initialized: ['score.weight']
+You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
+[2025-05-26 15:57:36,548][__main__][INFO] - Loaded pre-trained PEFT model from kamel-usp/jbcs2025_phi4-balanced-C4
+[2025-05-26 15:57:36,553][__main__][INFO] - None
+[2025-05-26 15:57:36,558][transformers.training_args][INFO] - PyTorch: setting up devices
+[2025-05-26 15:57:36,657][transformers.training_args][INFO] - The default value for the training argument `--report_to` will change in v5 (from all installed integrations to none). In v5, you will need to use `--report_to all` to get the same behavior as now. You should start updating your code and make this info disappear :-).
+[2025-05-26 15:57:36,678][transformers.trainer][INFO] - You have loaded a model on multiple GPUs. `is_model_parallel` attribute will be force-set to `True` to avoid any unexpected behavior such as device placement mismatching.
+[2025-05-26 15:57:36,719][transformers.trainer][INFO] - Using auto half precision backend
+[2025-05-26 15:57:36,720][transformers.trainer][WARNING] - No label_names provided for model class `PeftModelForSequenceClassification`. Since `PeftModel` hides base models input arguments, if label_names is not given, label_names can't be set automatically within `Trainer`. Note that empty label_names list will be used instead.
+[2025-05-26 15:57:36,720][__main__][INFO] - Running inference on test dataset
+[2025-05-26 15:57:36,722][transformers.trainer][INFO] - The following columns in the test set don't have a corresponding argument in `PeftModelForSequenceClassification.forward` and have been ignored: id, grades, prompt, essay_year, supporting_text, essay_text, id_prompt, reference. If id, grades, prompt, essay_year, supporting_text, essay_text, id_prompt, reference are not expected by `PeftModelForSequenceClassification.forward`,  you can safely ignore this message.
+[2025-05-26 15:57:36,754][transformers.trainer][INFO] - 
+***** Running Prediction *****
+[2025-05-26 15:57:36,754][transformers.trainer][INFO] -   Num examples = 138
+[2025-05-26 15:57:36,754][transformers.trainer][INFO] -   Batch size = 4
+[2025-05-26 15:59:37,941][transformers][INFO] - {'accuracy': 0.7028985507246377, 'RMSE': 24.55399256179405, 'QWK': 0.579465541490858, 'HDIV': 0.007246376811594235, 'Macro_F1': 0.27468966776195697, 'Micro_F1': 0.7028985507246377, 'Weighted_F1': 0.6761155293109196, 'TP_0': np.int64(0), 'TN_0': np.int64(137), 'FP_0': np.int64(0), 'FN_0': np.int64(1), 'TP_1': np.int64(0), 'TN_1': np.int64(135), 'FP_1': np.int64(2), 'FN_1': np.int64(1), 'TP_2': np.int64(1), 'TN_2': np.int64(128), 'FP_2': np.int64(1), 'FN_2': np.int64(8), 'TP_3': np.int64(67), 'TN_3': np.int64(39), 'FP_3': np.int64(23), 'FN_3': np.int64(9), 'TP_4': np.int64(29), 'TN_4': np.int64(79), 'FP_4': np.int64(13), 'FN_4': np.int64(17), 'TP_5': np.int64(0), 'TN_5': np.int64(131), 'FP_5': np.int64(2), 'FN_5': np.int64(5)}
+[2025-05-26 15:59:37,967][__main__][INFO] - Inference results saved to phi-4-phi4_classification_lora-C4_inference_results.jsonl
+[2025-05-26 15:59:37,967][__main__][INFO] - Inference results: {'accuracy': 0.7028985507246377, 'RMSE': 24.55399256179405, 'QWK': 0.579465541490858, 'HDIV': 0.007246376811594235, 'Macro_F1': 0.27468966776195697, 'Micro_F1': 0.7028985507246377, 'Weighted_F1': 0.6761155293109196, 'TP_0': np.int64(0), 'TN_0': np.int64(137), 'FP_0': np.int64(0), 'FN_0': np.int64(1), 'TP_1': np.int64(0), 'TN_1': np.int64(135), 'FP_1': np.int64(2), 'FN_1': np.int64(1), 'TP_2': np.int64(1), 'TN_2': np.int64(128), 'FP_2': np.int64(1), 'FN_2': np.int64(8), 'TP_3': np.int64(67), 'TN_3': np.int64(39), 'FP_3': np.int64(23), 'FN_3': np.int64(9), 'TP_4': np.int64(29), 'TN_4': np.int64(79), 'FP_4': np.int64(13), 'FN_4': np.int64(17), 'TP_5': np.int64(0), 'TN_5': np.int64(131), 'FP_5': np.int64(2), 'FN_5': np.int64(5)}
+[2025-05-26 15:59:37,968][__main__][INFO] - Inference experiment completed

runs/slm_decoder_models/phi-4/phi-4-phi4_classification_lora-C5/.hydra/config.yaml

@@ -0,0 +1,37 @@
+cache_dir: /tmp/
+dataset:
+  name: kamel-usp/aes_enem_dataset
+  split: JBCS2025
+training_params:
+  seed: 42
+  num_train_epochs: 20
+  logging_steps: 100
+  metric_for_best_model: QWK
+  bf16: true
+post_training_results:
+  model_path: /workspace/jbcs2025/outputs/2025-03-24/20-42-59
+experiments:
+  model:
+    name: microsoft/phi-4
+    type: phi4_classification_lora
+    num_labels: 6
+    output_dir: ./results/phi4-balanced/C5
+    logging_dir: ./logs/phi4-balanced/C5
+    best_model_dir: ./results/phi4-balanced/C5/best_model
+    lora_r: 8
+    lora_dropout: 0.05
+    lora_alpha: 16
+    lora_target_modules: all-linear
+    checkpoint_path: kamel-usp/jbcs2025_phi4-balanced-C5
+  tokenizer:
+    name: microsoft/phi-4
+  dataset:
+    grade_index: 4
+  training_params:
+    weight_decay: 0.01
+    warmup_ratio: 0.1
+    learning_rate: 5.0e-05
+    train_batch_size: 2
+    eval_batch_size: 4
+    gradient_accumulation_steps: 8
+    gradient_checkpointing: false

runs/slm_decoder_models/phi-4/phi-4-phi4_classification_lora-C5/.hydra/hydra.yaml

@@ -0,0 +1,155 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: run_inference_experiment
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: config
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.1'
+    cwd: /workspace/jbcs2025
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /workspace/jbcs2025/configs
+      schema: file
+      provider: main
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /workspace/jbcs2025/outputs/2025-05-26/16-00-16
+    choices:
+      experiments: slm_decoder_models/C5
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false

runs/slm_decoder_models/phi-4/phi-4-phi4_classification_lora-C5/.hydra/overrides.yaml

@@ -0,0 +1 @@
+[]

runs/slm_decoder_models/phi-4/phi-4-phi4_classification_lora-C5/evaluation_results.csv

@@ -0,0 +1,2 @@
+accuracy,RMSE,QWK,HDIV,Macro_F1,Micro_F1,Weighted_F1,TP_0,TN_0,FP_0,FN_0,TP_1,TN_1,FP_1,FN_1,TP_2,TN_2,FP_2,FN_2,TP_3,TN_3,FP_3,FN_3,TP_4,TN_4,FP_4,FN_4,TP_5,TN_5,FP_5,FN_5,timestamp,id
+0.3115942028985507,60.91095901015048,0.45744053469628465,0.1376811594202898,0.18965878221692178,0.3115942028985507,0.24393006835878425,3,115,1,19,17,55,51,15,1,103,11,23,0,113,0,25,22,74,32,10,0,135,0,3,2025-05-26 16:00:16,phi-4-phi4_classification_lora-C5

runs/slm_decoder_models/phi-4/phi-4-phi4_classification_lora-C5/run_inference_experiment.log

@@ -0,0 +1,124 @@
+[2025-05-26 16:00:16,898][__main__][INFO] - Starting inference experiment
+[2025-05-26 16:00:16,900][__main__][INFO] - cache_dir: /tmp/
+dataset:
+  name: kamel-usp/aes_enem_dataset
+  split: JBCS2025
+training_params:
+  seed: 42
+  num_train_epochs: 20
+  logging_steps: 100
+  metric_for_best_model: QWK
+  bf16: true
+post_training_results:
+  model_path: /workspace/jbcs2025/outputs/2025-03-24/20-42-59
+experiments:
+  model:
+    name: microsoft/phi-4
+    type: phi4_classification_lora
+    num_labels: 6
+    output_dir: ./results/phi4-balanced/C5
+    logging_dir: ./logs/phi4-balanced/C5
+    best_model_dir: ./results/phi4-balanced/C5/best_model
+    lora_r: 8
+    lora_dropout: 0.05
+    lora_alpha: 16
+    lora_target_modules: all-linear
+    checkpoint_path: kamel-usp/jbcs2025_phi4-balanced-C5
+  tokenizer:
+    name: microsoft/phi-4
+  dataset:
+    grade_index: 4
+  training_params:
+    weight_decay: 0.01
+    warmup_ratio: 0.1
+    learning_rate: 5.0e-05
+    train_batch_size: 2
+    eval_batch_size: 4
+    gradient_accumulation_steps: 8
+    gradient_checkpointing: false
+
+[2025-05-26 16:00:16,903][__main__][INFO] - Running inference with fine-tuned HF model
+[2025-05-26 16:00:22,843][transformers.tokenization_utils_base][INFO] - loading file vocab.json from cache at /tmp/models--microsoft--phi-4/snapshots/187ef0342fff0eb3333be9f00389385e95ef0b61/vocab.json
+[2025-05-26 16:00:22,843][transformers.tokenization_utils_base][INFO] - loading file merges.txt from cache at /tmp/models--microsoft--phi-4/snapshots/187ef0342fff0eb3333be9f00389385e95ef0b61/merges.txt
+[2025-05-26 16:00:22,843][transformers.tokenization_utils_base][INFO] - loading file tokenizer.json from cache at /tmp/models--microsoft--phi-4/snapshots/187ef0342fff0eb3333be9f00389385e95ef0b61/tokenizer.json
+[2025-05-26 16:00:22,843][transformers.tokenization_utils_base][INFO] - loading file added_tokens.json from cache at /tmp/models--microsoft--phi-4/snapshots/187ef0342fff0eb3333be9f00389385e95ef0b61/added_tokens.json
+[2025-05-26 16:00:22,844][transformers.tokenization_utils_base][INFO] - loading file special_tokens_map.json from cache at /tmp/models--microsoft--phi-4/snapshots/187ef0342fff0eb3333be9f00389385e95ef0b61/special_tokens_map.json
+[2025-05-26 16:00:22,844][transformers.tokenization_utils_base][INFO] - loading file tokenizer_config.json from cache at /tmp/models--microsoft--phi-4/snapshots/187ef0342fff0eb3333be9f00389385e95ef0b61/tokenizer_config.json
+[2025-05-26 16:00:22,844][transformers.tokenization_utils_base][INFO] - loading file chat_template.jinja from cache at None
+[2025-05-26 16:00:23,193][__main__][INFO] - Tokenizer function parameters- Padding:longest; Truncation: False
+[2025-05-26 16:00:25,150][__main__][INFO] - Loading model from: microsoft/phi-4
+[2025-05-26 16:00:25,402][transformers.configuration_utils][INFO] - loading configuration file config.json from cache at /tmp/models--microsoft--phi-4/snapshots/187ef0342fff0eb3333be9f00389385e95ef0b61/config.json
+[2025-05-26 16:00:25,405][transformers.configuration_utils][INFO] - Model config Phi3Config {
+  "architectures": [
+    "Phi3ForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "bos_token_id": 100257,
+  "embd_pdrop": 0.0,
+  "eos_token_id": 100265,
+  "hidden_act": "silu",
+  "hidden_size": 5120,
+  "id2label": {
+    "0": 0,
+    "1": 40,
+    "2": 80,
+    "3": 120,
+    "4": 160,
+    "5": 200
+  },
+  "initializer_range": 0.02,
+  "intermediate_size": 17920,
+  "label2id": {
+    "0": 0,
+    "40": 1,
+    "80": 2,
+    "120": 3,
+    "160": 4,
+    "200": 5
+  },
+  "max_position_embeddings": 16384,
+  "model_type": "phi3",
+  "num_attention_heads": 40,
+  "num_hidden_layers": 40,
+  "num_key_value_heads": 10,
+  "original_max_position_embeddings": 16384,
+  "pad_token_id": 100349,
+  "partial_rotary_factor": 1.0,
+  "resid_pdrop": 0.0,
+  "rms_norm_eps": 1e-05,
+  "rope_scaling": null,
+  "rope_theta": 250000,
+  "sliding_window": null,
+  "tie_word_embeddings": false,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.52.3",
+  "use_cache": true,
+  "vocab_size": 100352
+}
+
+[2025-05-26 16:00:25,642][transformers.modeling_utils][INFO] - loading weights file model.safetensors from cache at /tmp/models--microsoft--phi-4/snapshots/187ef0342fff0eb3333be9f00389385e95ef0b61/model.safetensors.index.json
+[2025-05-26 16:00:25,642][transformers.modeling_utils][INFO] - Will use torch_dtype=torch.bfloat16 as defined in model's config object
+[2025-05-26 16:00:25,642][transformers.modeling_utils][INFO] - Instantiating Phi3ForSequenceClassification model under default dtype torch.bfloat16.
+[2025-05-26 16:00:34,576][transformers.modeling_utils][INFO] - Some weights of the model checkpoint at microsoft/phi-4 were not used when initializing Phi3ForSequenceClassification: ['lm_head.weight']
+- This IS expected if you are initializing Phi3ForSequenceClassification from the checkpoint of a model trained on another task or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPreTraining model).
+- This IS NOT expected if you are initializing Phi3ForSequenceClassification from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).
+[2025-05-26 16:00:34,576][transformers.modeling_utils][WARNING] - Some weights of Phi3ForSequenceClassification were not initialized from the model checkpoint at microsoft/phi-4 and are newly initialized: ['score.weight']
+You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
+[2025-05-26 16:00:42,106][__main__][INFO] - Loaded pre-trained PEFT model from kamel-usp/jbcs2025_phi4-balanced-C5
+[2025-05-26 16:00:42,112][__main__][INFO] - None
+[2025-05-26 16:00:42,117][transformers.training_args][INFO] - PyTorch: setting up devices
+[2025-05-26 16:00:42,170][transformers.training_args][INFO] - The default value for the training argument `--report_to` will change in v5 (from all installed integrations to none). In v5, you will need to use `--report_to all` to get the same behavior as now. You should start updating your code and make this info disappear :-).
+[2025-05-26 16:00:42,190][transformers.trainer][INFO] - You have loaded a model on multiple GPUs. `is_model_parallel` attribute will be force-set to `True` to avoid any unexpected behavior such as device placement mismatching.
+[2025-05-26 16:00:42,228][transformers.trainer][INFO] - Using auto half precision backend
+[2025-05-26 16:00:42,229][transformers.trainer][WARNING] - No label_names provided for model class `PeftModelForSequenceClassification`. Since `PeftModel` hides base models input arguments, if label_names is not given, label_names can't be set automatically within `Trainer`. Note that empty label_names list will be used instead.
+[2025-05-26 16:00:42,230][__main__][INFO] - Running inference on test dataset
+[2025-05-26 16:00:42,231][transformers.trainer][INFO] - The following columns in the test set don't have a corresponding argument in `PeftModelForSequenceClassification.forward` and have been ignored: id, id_prompt, essay_text, essay_year, grades, supporting_text, reference, prompt. If id, id_prompt, essay_text, essay_year, grades, supporting_text, reference, prompt are not expected by `PeftModelForSequenceClassification.forward`,  you can safely ignore this message.
+[2025-05-26 16:00:42,262][transformers.trainer][INFO] - 
+***** Running Prediction *****
+[2025-05-26 16:00:42,262][transformers.trainer][INFO] -   Num examples = 138
+[2025-05-26 16:00:42,262][transformers.trainer][INFO] -   Batch size = 4
+[2025-05-26 16:02:50,833][transformers][INFO] - {'accuracy': 0.3115942028985507, 'RMSE': 60.91095901015048, 'QWK': 0.45744053469628465, 'HDIV': 0.1376811594202898, 'Macro_F1': 0.18965878221692178, 'Micro_F1': 0.3115942028985507, 'Weighted_F1': 0.24393006835878425, 'TP_0': np.int64(3), 'TN_0': np.int64(115), 'FP_0': np.int64(1), 'FN_0': np.int64(19), 'TP_1': np.int64(17), 'TN_1': np.int64(55), 'FP_1': np.int64(51), 'FN_1': np.int64(15), 'TP_2': np.int64(1), 'TN_2': np.int64(103), 'FP_2': np.int64(11), 'FN_2': np.int64(23), 'TP_3': np.int64(0), 'TN_3': np.int64(113), 'FP_3': np.int64(0), 'FN_3': np.int64(25), 'TP_4': np.int64(22), 'TN_4': np.int64(74), 'FP_4': np.int64(32), 'FN_4': np.int64(10), 'TP_5': np.int64(0), 'TN_5': np.int64(135), 'FP_5': np.int64(0), 'FN_5': np.int64(3)}
+[2025-05-26 16:02:50,854][__main__][INFO] - Inference results saved to phi-4-phi4_classification_lora-C5_inference_results.jsonl
+[2025-05-26 16:02:50,855][__main__][INFO] - Inference results: {'accuracy': 0.3115942028985507, 'RMSE': 60.91095901015048, 'QWK': 0.45744053469628465, 'HDIV': 0.1376811594202898, 'Macro_F1': 0.18965878221692178, 'Micro_F1': 0.3115942028985507, 'Weighted_F1': 0.24393006835878425, 'TP_0': np.int64(3), 'TN_0': np.int64(115), 'FP_0': np.int64(1), 'FN_0': np.int64(19), 'TP_1': np.int64(17), 'TN_1': np.int64(55), 'FP_1': np.int64(51), 'FN_1': np.int64(15), 'TP_2': np.int64(1), 'TN_2': np.int64(103), 'FP_2': np.int64(11), 'FN_2': np.int64(23), 'TP_3': np.int64(0), 'TN_3': np.int64(113), 'FP_3': np.int64(0), 'FN_3': np.int64(25), 'TP_4': np.int64(22), 'TN_4': np.int64(74), 'FP_4': np.int64(32), 'FN_4': np.int64(10), 'TP_5': np.int64(0), 'TN_5': np.int64(135), 'FP_5': np.int64(0), 'FN_5': np.int64(3)}
+[2025-05-26 16:02:50,855][__main__][INFO] - Inference experiment completed