opensportslib 0.2.0.dev2__tar.gz → 0.2.0.dev3__tar.gz

{opensportslib-0.2.0.dev2/opensportslib.egg-info → opensportslib-0.2.0.dev3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: opensportslib
-Version: 0.2.0.dev2
+Version: 0.2.0.dev3
 Summary: OpenSportsLib is the professional library, designed for advanced video understanding in sports. It provides state-of-the-art tools for action recognition, spotting, retrieval, and captioning, making it ideal for researchers, analysts, and developers working with sports video data.
 Author: Jeet Vora
 Requires-Python: >=3.12
@@ -93,13 +93,26 @@ opensportslib setup --pyg
 
 # Optional: install for DALI support
 opensportslib setup --dali
-``` 
+
+# Optional: install the X-VARS-compatible VQA dependency profile
+opensportslib setup --vqa_xvars
+
+# Optional: install the Qwen-compatible VQA dependency profile
+opensportslib setup --vqa_qwen
+```
 ---
 
 **Note:**  
 Run `opensportslib setup` to automatically configure dependencies.  
 If issues occur, manually install compatible versions of `torch`, `torchvision`, and related libraries according to your CUDA version or system compatibility.
 
+For VQA, use exactly one backend-specific dependency profile:
+
+- `--vqa_xvars` installs the X-VARS-compatible Hugging Face stack from `XVARS_DEPENDENCY_PINS`
+- `--vqa_qwen` installs the Qwen-compatible Hugging Face stack from `QWEN_DEPENDENCY_PINS`
+
+The `vqa_qwen` config supports `Qwen/Qwen2.5-7B-Instruct` and `Qwen/Qwen3.5-9B-Base`.
+
 ---
 
 ## Data and pretrained models
@@ -287,7 +300,7 @@ metrics_from_file = my_model.evaluate(
 from opensportslib.apis import VQAModel
 
 my_model = VQAModel(
-    config="/path/to/vqa.yaml",
+    config="opensportslib/configs/vqa/qwen.yaml",
     weights=None,  # optional: path or Hugging Face model ID
 )
 
@@ -300,13 +313,20 @@ single_prediction = my_model.infer(
     video_path="/path/to/video.mp4",
     question="What card would you give? Why?",
 )
-
-metrics = my_model.evaluate(
-    test_set="/path/to/test_annotations.json",
-    predictions=predictions,
-)
 ```
 
+Use `opensportslib/configs/vqa/xvars.yaml` with `opensportslib setup --vqa_xvars`
+for the X-VARS-compatible backend, or `opensportslib/configs/vqa/qwen.yaml` with
+`opensportslib setup --vqa_qwen` for the Qwen-compatible backend. The Qwen
+backend currently supports `Qwen/Qwen2.5-7B-Instruct` and
+`Qwen/Qwen3.5-9B-Base`.
+
+For X-VARS, `feature_source: indexed_or_raw_clip` prefers indexed CLIP features
+when available and falls back to extracting CLIP features from raw video during
+`infer()`. Pre-extracted features remain the preferred path for parity, speed,
+and reproducibility. See [docs/xvars_integration_phases.md](docs/xvars_integration_phases.md)
+for the full X-VARS setup workflow.
+
 
 ---
 
@@ -414,6 +434,12 @@ opensportslib setup --pyg
 
 # Optional: install for DALI support
 opensportslib setup --dali
+
+# Optional: install the X-VARS-compatible VQA dependency profile
+opensportslib setup --vqa_xvars
+
+# Optional: install the Qwen-compatible VQA dependency profile
+opensportslib setup --vqa_qwen
 ```
 
 ### Git workflow

{opensportslib-0.2.0.dev2 → opensportslib-0.2.0.dev3}/README.md

@@ -58,13 +58,26 @@ opensportslib setup --pyg
 
 # Optional: install for DALI support
 opensportslib setup --dali
-``` 
+
+# Optional: install the X-VARS-compatible VQA dependency profile
+opensportslib setup --vqa_xvars
+
+# Optional: install the Qwen-compatible VQA dependency profile
+opensportslib setup --vqa_qwen
+```
 ---
 
 **Note:**  
 Run `opensportslib setup` to automatically configure dependencies.  
 If issues occur, manually install compatible versions of `torch`, `torchvision`, and related libraries according to your CUDA version or system compatibility.
 
+For VQA, use exactly one backend-specific dependency profile:
+
+- `--vqa_xvars` installs the X-VARS-compatible Hugging Face stack from `XVARS_DEPENDENCY_PINS`
+- `--vqa_qwen` installs the Qwen-compatible Hugging Face stack from `QWEN_DEPENDENCY_PINS`
+
+The `vqa_qwen` config supports `Qwen/Qwen2.5-7B-Instruct` and `Qwen/Qwen3.5-9B-Base`.
+
 ---
 
 ## Data and pretrained models
@@ -252,7 +265,7 @@ metrics_from_file = my_model.evaluate(
 from opensportslib.apis import VQAModel
 
 my_model = VQAModel(
-    config="/path/to/vqa.yaml",
+    config="opensportslib/configs/vqa/qwen.yaml",
     weights=None,  # optional: path or Hugging Face model ID
 )
 
@@ -265,13 +278,20 @@ single_prediction = my_model.infer(
     video_path="/path/to/video.mp4",
     question="What card would you give? Why?",
 )
-
-metrics = my_model.evaluate(
-    test_set="/path/to/test_annotations.json",
-    predictions=predictions,
-)
 ```
 
+Use `opensportslib/configs/vqa/xvars.yaml` with `opensportslib setup --vqa_xvars`
+for the X-VARS-compatible backend, or `opensportslib/configs/vqa/qwen.yaml` with
+`opensportslib setup --vqa_qwen` for the Qwen-compatible backend. The Qwen
+backend currently supports `Qwen/Qwen2.5-7B-Instruct` and
+`Qwen/Qwen3.5-9B-Base`.
+
+For X-VARS, `feature_source: indexed_or_raw_clip` prefers indexed CLIP features
+when available and falls back to extracting CLIP features from raw video during
+`infer()`. Pre-extracted features remain the preferred path for parity, speed,
+and reproducibility. See [docs/xvars_integration_phases.md](docs/xvars_integration_phases.md)
+for the full X-VARS setup workflow.
+
 
 ---
 
@@ -379,6 +399,12 @@ opensportslib setup --pyg
 
 # Optional: install for DALI support
 opensportslib setup --dali
+
+# Optional: install the X-VARS-compatible VQA dependency profile
+opensportslib setup --vqa_xvars
+
+# Optional: install the Qwen-compatible VQA dependency profile
+opensportslib setup --vqa_qwen
 ```
 
 ### Git workflow

{opensportslib-0.2.0.dev2 → opensportslib-0.2.0.dev3}/opensportslib/cli.py

@@ -11,8 +11,8 @@ def main(argv: Optional[list[str]] = None) -> int:
     parser.add_argument("command", choices=["setup"])
     parser.add_argument("--pyg", action="store_true")
     parser.add_argument("--dali", action="store_true")
-    parser.add_argument("--xvars", action="store_true")
-    parser.add_argument("--qwen", action="store_true")
+    parser.add_argument("--vqa_xvars", action="store_true")
+    parser.add_argument("--vqa_qwen", action="store_true")
 
     args = parser.parse_args(argv)
 
@@ -20,8 +20,8 @@ def main(argv: Optional[list[str]] = None) -> int:
         setup(
             pyg=args.pyg,
             dali=args.dali,
-            xvars=args.xvars,
-            qwen=args.qwen
+            vqa_xvars=args.vqa_xvars,
+            vqa_qwen=args.vqa_qwen
         )
         return 0
 

opensportslib-0.2.0.dev3/opensportslib/configs/vqa/qwen.yaml

@@ -0,0 +1,49 @@
+SYSTEM:
+  paths:
+    save_dir: ./checkpoints_vqa_qwen
+
+MODEL:
+  components:
+    video_encoder:
+      kind: encoder
+      source:
+        provider: opensportslib
+        name: xvars_clip_features
+      load:
+        weights_path: /home/vorajv/X-VARS/weights/14_model.pth.tar
+      params:
+        feature_source: indexed_or_raw_clip
+        vision_tower: openai/clip-vit-large-patch14
+        feature_dim: 1024
+      overrides: {}
+    mm_projector:
+      kind: projector
+      source:
+        provider: opensportslib
+      params:
+        input_dim: 1024
+      overrides: {}
+    llm_decoder:
+      kind: decoder
+      source:
+        provider: huggingface
+        # Supported models: Qwen/Qwen2.5-7B-Instruct or Qwen/Qwen3.5-9B-Base
+        name: Qwen/Qwen3.5-9B-Base
+      params:
+        # Supported models: Qwen/Qwen2.5-7B-Instruct or Qwen/Qwen3.5-9B-Base
+        repo_id: Qwen/Qwen3.5-9B-Base
+      overrides: {}
+  metadata:
+    backend: qwen_xvars_infer
+
+TRAIN:
+  execution:
+    prompt:
+      system_prompt: You are a football video assistant. Answer the VQA question using the provided video context and referee priors.
+
+    hf:
+      tokenizer_id: ${MODEL.components.llm_decoder.params.repo_id}
+      prefer_cuda: true
+      local_files_only: false
+      device_map: auto
+      offload_folder: ./hf_offload_qwen

{opensportslib-0.2.0.dev2 → opensportslib-0.2.0.dev3}/opensportslib/core/config/loader.py

@@ -18,7 +18,7 @@ from .migrate import migrate_config
 from .runtime_adapter import maybe_namespace, namespace_to_plain_dict
 
 _YAML_SUFFIXES = {".yaml", ".yml"}
-_TASK_DIRS = {"classification", "localization"}
+_TASK_DIRS = {"classification", "localization", "vqa"}
 _INTERPOLATION_RE = re.compile(r"\$\{([^}]+)\}")
 
 

{opensportslib-0.2.0.dev2 → opensportslib-0.2.0.dev3}/opensportslib/setup/setup.py

@@ -183,12 +183,12 @@ def verify():
     else:
         print("Running on CPU")
 
-def setup(dali=False, pyg=False, xvars=False, qwen=False):
+def setup(dali=False, pyg=False, vqa_xvars=False, vqa_qwen=False):
     install_torch()
     install_extras(dali=dali, pyg=pyg)
-    if xvars:
+    if vqa_xvars:
         install_xvars_dependencies(XVARS_DEPENDENCY_PINS)
-    if qwen:
+    if vqa_qwen:
         install_xvars_dependencies(QWEN_DEPENDENCY_PINS)
     verify()
 
@@ -202,9 +202,9 @@ if __name__ == "__main__":
     parser = argparse.ArgumentParser()
     parser.add_argument("--dali", action="store_true")
     parser.add_argument("--pyg", action="store_true")
-    parser.add_argument("--xvars", action="store_true")
-    parser.add_argument("--qwen", action="store_true")
+    parser.add_argument("--vqa_xvars", action="store_true")
+    parser.add_argument("--vqa_qwen", action="store_true")
 
     args = parser.parse_args()
 
-    setup(dali=args.dali, pyg=args.pyg, xvars=args.xvars, qwen=args.qwen)
+    setup(dali=args.dali, pyg=args.pyg, vqa_xvars=args.vqa_xvars, vqa_qwen=args.vqa_qwen)

{opensportslib-0.2.0.dev2 → opensportslib-0.2.0.dev3/opensportslib.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: opensportslib
-Version: 0.2.0.dev2
+Version: 0.2.0.dev3
 Summary: OpenSportsLib is the professional library, designed for advanced video understanding in sports. It provides state-of-the-art tools for action recognition, spotting, retrieval, and captioning, making it ideal for researchers, analysts, and developers working with sports video data.
 Author: Jeet Vora
 Requires-Python: >=3.12
@@ -93,13 +93,26 @@ opensportslib setup --pyg
 
 # Optional: install for DALI support
 opensportslib setup --dali
-``` 
+
+# Optional: install the X-VARS-compatible VQA dependency profile
+opensportslib setup --vqa_xvars
+
+# Optional: install the Qwen-compatible VQA dependency profile
+opensportslib setup --vqa_qwen
+```
 ---
 
 **Note:**  
 Run `opensportslib setup` to automatically configure dependencies.  
 If issues occur, manually install compatible versions of `torch`, `torchvision`, and related libraries according to your CUDA version or system compatibility.
 
+For VQA, use exactly one backend-specific dependency profile:
+
+- `--vqa_xvars` installs the X-VARS-compatible Hugging Face stack from `XVARS_DEPENDENCY_PINS`
+- `--vqa_qwen` installs the Qwen-compatible Hugging Face stack from `QWEN_DEPENDENCY_PINS`
+
+The `vqa_qwen` config supports `Qwen/Qwen2.5-7B-Instruct` and `Qwen/Qwen3.5-9B-Base`.
+
 ---
 
 ## Data and pretrained models
@@ -287,7 +300,7 @@ metrics_from_file = my_model.evaluate(
 from opensportslib.apis import VQAModel
 
 my_model = VQAModel(
-    config="/path/to/vqa.yaml",
+    config="opensportslib/configs/vqa/qwen.yaml",
     weights=None,  # optional: path or Hugging Face model ID
 )
 
@@ -300,13 +313,20 @@ single_prediction = my_model.infer(
     video_path="/path/to/video.mp4",
     question="What card would you give? Why?",
 )
-
-metrics = my_model.evaluate(
-    test_set="/path/to/test_annotations.json",
-    predictions=predictions,
-)
 ```
 
+Use `opensportslib/configs/vqa/xvars.yaml` with `opensportslib setup --vqa_xvars`
+for the X-VARS-compatible backend, or `opensportslib/configs/vqa/qwen.yaml` with
+`opensportslib setup --vqa_qwen` for the Qwen-compatible backend. The Qwen
+backend currently supports `Qwen/Qwen2.5-7B-Instruct` and
+`Qwen/Qwen3.5-9B-Base`.
+
+For X-VARS, `feature_source: indexed_or_raw_clip` prefers indexed CLIP features
+when available and falls back to extracting CLIP features from raw video during
+`infer()`. Pre-extracted features remain the preferred path for parity, speed,
+and reproducibility. See [docs/xvars_integration_phases.md](docs/xvars_integration_phases.md)
+for the full X-VARS setup workflow.
+
 
 ---
 
@@ -414,6 +434,12 @@ opensportslib setup --pyg
 
 # Optional: install for DALI support
 opensportslib setup --dali
+
+# Optional: install the X-VARS-compatible VQA dependency profile
+opensportslib setup --vqa_xvars
+
+# Optional: install the Qwen-compatible VQA dependency profile
+opensportslib setup --vqa_qwen
 ```
 
 ### Git workflow

{opensportslib-0.2.0.dev2 → opensportslib-0.2.0.dev3}/opensportslib.egg-info/SOURCES.txt

@@ -29,7 +29,6 @@ opensportslib/configs/localization/netvladpp_resnetpca512.yaml
 opensportslib/configs/localization/video_dali.yaml
 opensportslib/configs/localization/video_ocv.yaml
 opensportslib/configs/vqa/qwen.yaml
-opensportslib/configs/vqa/xvars_lora.yaml
 opensportslib/core/__init__.py
 opensportslib/core/config/__init__.py
 opensportslib/core/config/accessors.py

{opensportslib-0.2.0.dev2 → opensportslib-0.2.0.dev3}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "opensportslib"
-version = "0.2.0.dev2"
+version = "0.2.0.dev3"
 description = "OpenSportsLib is the professional library, designed for advanced video understanding in sports. It provides state-of-the-art tools for action recognition, spotting, retrieval, and captioning, making it ideal for researchers, analysts, and developers working with sports video data."
 readme = "README.md"
 requires-python = ">=3.12"

{opensportslib-0.2.0.dev2 → opensportslib-0.2.0.dev3}/tests/test_config_architecture.py

@@ -1,7 +1,6 @@
 from pathlib import Path
 
 import pytest
-import yaml
 
 from opensportslib.core.config import load_config, migrate_config, validate_config
 from opensportslib.core.config.accessors import (
@@ -25,16 +24,19 @@ def test_public_config_api_is_canonical_first():
 def test_legacy_inputs_route_through_migration(tmp_path):
     config_path = tmp_path / "legacy.yaml"
     config_path.write_text(
-        yaml.safe_dump(
-            {
-                "DATA": {
-                    "data_dir": str(tmp_path / "data"),
-                    "annotations": {"train": str(tmp_path / "train.json")},
-                },
-                "MODEL": {"backbone": {"type": "smoke_backbone"}},
-                "SYSTEM": {"save_dir": str(tmp_path / "ckpt")},
-            },
-            sort_keys=False,
+        "\n".join(
+            [
+                "DATA:",
+                f"  data_dir: {tmp_path / 'data'}",
+                "  annotations:",
+                f"    train: {tmp_path / 'train.json'}",
+                "MODEL:",
+                "  backbone:",
+                "    type: smoke_backbone",
+                "SYSTEM:",
+                f"  save_dir: {tmp_path / 'ckpt'}",
+                "",
+            ]
         ),
         encoding="utf-8",
     )
@@ -83,6 +85,33 @@ def test_localization_experiment_composes_all_layers():
     assert cfg["MODEL"]["components"]["video_encoder"]["source"]["name"] == "rny008_gsm"
 
 
+def test_vqa_xvars_experiment_composes_all_layers():
+    cfg = load_config("opensportslib/configs/vqa/xvars.yaml", as_namespace=False)
+
+    assert cfg["VERSION"] == 2
+    assert cfg["TASK"] == "vqa"
+    assert cfg["SYSTEM"]["paths"]["log_dir"] == "./logs"
+    assert cfg["SYSTEM"]["paths"]["save_dir"] == "./checkpoints_vqa_lora"
+    assert cfg["SYSTEM"]["paths"]["work_dir"] == "./checkpoints_vqa_lora"
+    assert cfg["DATA"]["common"]["runtime"]["loader_backend"] == "opencv"
+    assert cfg["MODEL"]["metadata"]["backend"] == "xvars_videochatgpt"
+    assert cfg["TRAIN"]["execution"]["hf"]["tokenizer_id"] == "/home/vorajv/X-VARS/weights/base_model_videoChatGPT"
+
+
+def test_vqa_qwen_experiment_composes_all_layers():
+    cfg = load_config("opensportslib/configs/vqa/qwen.yaml", as_namespace=False)
+
+    assert cfg["VERSION"] == 2
+    assert cfg["TASK"] == "vqa"
+    assert cfg["SYSTEM"]["paths"]["log_dir"] == "./logs"
+    assert cfg["SYSTEM"]["paths"]["save_dir"] == "./checkpoints_vqa_qwen"
+    assert cfg["SYSTEM"]["paths"]["work_dir"] == "./checkpoints_vqa_qwen"
+    assert cfg["DATA"]["common"]["runtime"]["loader_backend"] == "opencv"
+    assert cfg["MODEL"]["metadata"]["backend"] == "qwen_xvars_infer"
+    assert cfg["MODEL"]["components"]["llm_decoder"]["source"]["name"] == "Qwen/Qwen3.5-9B-Base"
+    assert cfg["TRAIN"]["execution"]["hf"]["offload_folder"] == "./hf_offload_qwen"
+
+
 def test_validation_accepts_canonical_schema():
     canonical = load_config(
         "opensportslib/configs/localization/default.yaml",

opensportslib-0.2.0.dev2/opensportslib/configs/vqa/qwen.yaml

@@ -1,228 +0,0 @@
-TASK: VQA
-VERSION: 2
-
-SYSTEM:
-  paths:
-    log_dir: ./logs
-    save_dir: ./checkpoints_vqa_qwen
-    work_dir: ./checkpoints_vqa_qwen
-  device: cuda
-  gpu:
-    count: 1
-    id: 0
-  reproducibility:
-    use_seed: true
-    seed: 42
-
-DATA:
-  common:
-    dataset_name: OSL-XFoul
-    data_root: /home/vorajv/dataset/OSL-XFoul
-    feature_index: /home/vorajv/dataset/OSL-XFoul/feature_index.json
-    prediction_index: /home/vorajv/dataset/OSL-XFoul/prediction_index.json
-    runtime:
-      loader_backend: opencv
-    splits:
-      train:
-        annotation_path: /home/vorajv/dataset/OSL-XFoul/train.json
-        source_path: /home/vorajv/dataset/OSL-XFoul
-        dataloader:
-          batch_size: 1
-          shuffle: true
-          num_workers: 0
-          pin_memory: false
-          mp_context: spawn
-          persistent_workers: false
-      valid:
-        annotation_path: /home/vorajv/dataset/OSL-XFoul/valid.json
-        source_path: /home/vorajv/dataset/OSL-XFoul
-        dataloader:
-          batch_size: 1
-          shuffle: false
-          num_workers: 0
-          pin_memory: false
-          mp_context: spawn
-          persistent_workers: false
-      test:
-        annotation_path: /home/vorajv/dataset/OSL-XFoul/test.json
-        source_path: /home/vorajv/dataset/OSL-XFoul
-        dataloader:
-          batch_size: 1
-          shuffle: false
-          num_workers: 0
-          pin_memory: false
-          mp_context: spawn
-          persistent_workers: false
-  inputs:
-    video:
-      modality: video
-      representation: raw
-      source:
-        format: mp4
-      sampling:
-        num_frames: 100
-        input_fps: 25
-        target_fps: 17
-        start_frame: 63
-        end_frame: 87
-      transform: {}
-      augmentations: {}
-      params: {}
-    question:
-      modality: text
-      representation: raw
-      source:
-        format: json
-      sampling: {}
-      transform: {}
-      augmentations: {}
-      params: {}
-
-MODEL:
-  runtime:
-    dtype: fp16
-    device: auto
-    compile: false
-    freeze: false
-  load:
-    checkpoint_path: null
-    pretrained: false
-    strict: true
-    map_location: null
-    format: auto
-  components:
-    video_encoder:
-      kind: encoder
-      source:
-        provider: opensportslib
-        name: xvars_clip_features
-      load:
-        weights_path: /home/vorajv/X-VARS/weights/14_model.pth.tar
-      params:
-        feature_source: indexed_or_raw_clip
-        vision_tower: openai/clip-vit-large-patch14
-        feature_dim: 1024
-      overrides: {}
-    mm_projector:
-      kind: projector
-      source:
-        provider: opensportslib
-      params:
-        input_dim: 1024
-      overrides: {}
-    llm_decoder:
-      kind: decoder
-      source:
-        provider: huggingface
-        # Qwen/Qwen2.5-7B-Instruct or Qwen/Qwen3.5-9B-Base 
-        name: Qwen/Qwen3.5-9B-Base
-      params:
-        # Qwen/Qwen2.5-7B-Instruct or Qwen/Qwen3.5-9B-Base        
-        repo_id: Qwen/Qwen3.5-9B-Base
-      overrides: {}
-  topology:
-    - from: video_encoder
-      to: mm_projector
-    - from: mm_projector
-      to: llm_decoder
-  metadata:
-    backend: qwen_xvars_infer
-
-IO:
-  inputs:
-    video: video_encoder
-    question: llm_decoder
-  outputs:
-    answer_text: llm_decoder
-    explanation_text: llm_decoder
-
-TRAIN:
-  trainer:
-    type: vqa
-
-  epochs: 3
-
-  criterion:
-    type: CrossEntropyLoss
-
-  optimizer:
-    type: AdamW
-    lr: 0.0002
-    weight_decay: 0.001
-
-  scheduler:
-    type: constant
-
-  execution:
-    enabled: true
-    training_backend: xvars_videochatgpt_lora
-    feature_backend: xvars_clip
-    view_sampling_policy: random_train_deterministic_eval
-    acc_grad_iter: 8
-    log_interval: 1
-    dry_run: false
-
-    xvars:
-      feature_mode: strict_xvars
-      projection_path: null
-
-    prompt:
-      style: detailed
-      system_prompt: You are a football video assistant. Answer the VQA question using the provided video context and referee priors.
-      include_priors: true
-      prediction_prior_adapter: xvars_referee
-      prior_fields: [action, offence, contact, bodypart]
-      video_token_len: 300
-
-    generation:
-      max_new_tokens: 128
-      temperature: 0.0
-
-    eval_profile:
-      metric_set: [exact_match, contains_match, token_f1, referee_semantic]
-      aggregation: mean
-      exclusions: []
-
-    sft:
-      max_seq_length: 480
-      include_video_tokens: true
-      disable_tqdm: false
-      reference_mode: all
-      append_eos_token: true
-      gradient_checkpointing: true
-      save_strategy: epoch
-
-    hf:
-      tokenizer_id: ${MODEL.components.llm_decoder.params.repo_id}
-      prefer_cuda: true
-      local_files_only: false
-      device_map: auto
-      offload_folder: ./hf_offload_qwen
-
-    lora:
-      r: 16
-      alpha: 32
-      dropout: 0.05
-      bias: none
-      prepare_kbit: true
-      target_modules: [mm_projector, upsample_features, up_proj, down_proj, gate_proj, k_proj, q_proj, v_proj, o_proj]
-      exclude_modules: '^base_lm\.model\.mm_projector$'
-
-    quantization:
-      enabled: false
-      load_in_4bit: true
-      bnb_4bit_quant_type: nf4
-      compute_dtype: float16
-      bnb_4bit_use_double_quant: true
-
-    checkpoint:
-      save_adapter: true
-      merge_and_save: false
-
-  selection:
-    monitor: loss
-    mode: min
-
-  checkpoint:
-    save_every: 1
-    save_best: true