Getting ready to launch a new training run

pdelfin/train/config/qwen2vl-7b-lora.yaml

@@ -28,21 +28,20 @@ generate:
 train_data:
   seed: 1337
   sources:
-    - name: openai_batch_data_v2
-      query_glob_path: s3://ai2-oe-data/jakep/openai_batch_data_v2/*.jsonl
-      response_glob_path: s3://ai2-oe-data/jakep/openai_batch_done_v2/*.json
-      backend:
-        - openai
-      size: 100_000
+    - name: openai_batch_data_v5_1_eval # TODO This is just for testing the job, once ready change to a real train dataset
+      query_glob_path: s3://ai2-oe-data/jakep/pdfdata/openai_batch_data_v5_1_eval/*.jsonl
+      response_glob_path: s3://ai2-oe-data/jakep/pdfdata/openai_batch_done_v5_1_eval/*.jsonl
 
 valid_data:
   sources:
-    - name: openai_batch_data_eval_mini
-      query_glob_path: s3://ai2-oe-data/jakep/openai_batch_data_eval_mini/*.jsonl
-      response_glob_path: s3://ai2-oe-data/jakep/openai_batch_done_eval_mini/*.json
-      backend:
-        - openai
-      size: 100_000
+    - name: openai_batch_data_v5_1_eval
+      query_glob_path: s3://ai2-oe-data/jakep/pdfdata/openai_batch_data_v5_1_eval/*.jsonl
+      response_glob_path: s3://ai2-oe-data/jakep/pdfdata/openai_batch_done_v5_1_eval/*.jsonl
+    - name: openai_batch_data_v5_1_iabooks_eval
+      query_glob_path: s3://ai2-oe-data/jakep/pdfdata/openai_batch_data_v5_1_iabooks_eval/*.jsonl
+      response_glob_path: s3://ai2-oe-data/jakep/pdfdata/openai_batch_done_v5_1_iabooks_eval/*.jsonl
+
+
 
 # Mostly pulled from https://github.com/QwenLM/Qwen2/blob/main/examples/sft/finetune.sh
 hparams:
@@ -52,10 +51,10 @@ hparams:
   gradient_checkpointing: false
   clip_grad_norm: 1.0
   learning_rate: 3e-4
-  max_steps: 5000
+  max_steps: 2000
   pad_multiple_of: 16
   log_every_steps: 50
-  eval_every_steps: 500
+  eval_every_steps: 100
   optim: adamw_torch
   lr_scheduler: cosine
   weight_decay: 0.01

pdelfin/train/core/config.py

@@ -75,10 +75,8 @@ class AwsConfig:
 @dataclass
 class SourceConfig:
     name: str = field(help="The name of the source")
-    size: int = field(help="Limit size for the source")
     query_glob_path: str = field(help="The s3 bucket pointing to the inputs sent to OpenAI to generate the silver data")
     response_glob_path: str = field(help="The s3 bucket pointing to the batch api response json's sent back from open ai")
-    backend: List[str] = field(help="The data generation backend to use to train the model")
 
 
 @dataclass

pdelfin/train/train.py

@@ -15,7 +15,7 @@
 import accelerate
 import torch
 import torch.distributed
-from datasets import DatasetDict
+from datasets import DatasetDict, concatenate_datasets
 from datasets.utils import disable_progress_bars
 from datasets.utils.logging import set_verbosity
 from peft import LoraConfig, get_peft_model  # pyright: ignore
@@ -49,7 +49,7 @@
 )
 
 
-from pdelfin.train.dataloader import make_dataset
+from pdelfin.train.dataloader import build_batch_query_response_vision_dataset
 from pdelfin.train.dataprep import batch_prepare_data_for_qwen2_training, filter_by_max_seq_len
 
 
@@ -113,13 +113,6 @@ def run_train(config: TrainConfig):
 
     setup_environment(aws_config=config.aws, wandb_config=config.wandb, WANDB_RUN_GROUP=run_name.group)
 
-    dataset = make_dataset(
-        train_data_config=config.train_data,
-        valid_data_config=config.valid_data,
-        num_proc=config.num_proc,
-        logger=logger,
-    )
-
     model = Qwen2VLForConditionalGeneration.from_pretrained(
         config.model.name_or_path, torch_dtype=torch.bfloat16,
         _attn_implementation="flash_attention_2" if config.model.use_flash_attn else None
@@ -139,15 +132,33 @@ def run_train(config: TrainConfig):
         log_trainable_parameters(model=model, logger=logger)
 
     # Do final filtering, and prep for running model forward()
-    filtered_dataset = DatasetDict(**{split: dataset[split].filter(partial(filter_by_max_seq_len, processor=processor)) for split in dataset})
-    formatted_dataset = filtered_dataset.with_transform(partial(batch_prepare_data_for_qwen2_training, processor=processor))
-    print(formatted_dataset)
-    print("---------------")
-
+
+    # Training sets get all concatenated and shuffled
+    train_dataset = (
+        concatenate_datasets(
+            [
+                build_batch_query_response_vision_dataset(source.query_glob_path, source.response_glob_path)
+                for source in config.train_data.sources
+            ]
+        )
+        .filter(partial(filter_by_max_seq_len, processor=processor))
+        .with_transform(partial(batch_prepare_data_for_qwen2_training, processor=processor))
+    )
+
+    # Validation sets get put into a datasetdict so each can report a loss separately
+    valid_dataset = DatasetDict(
+        **{
+            source.name: build_batch_query_response_vision_dataset(source.query_glob_path, source.response_glob_path)
+            .filter(partial(filter_by_max_seq_len, processor=processor))
+            .with_transform(partial(batch_prepare_data_for_qwen2_training, processor=processor))
+            for source in config.valid_data.sources
+        }
+    )
+
     save_path = join_path("", config.save.path, run_name.run)
 
     save_config(config, join_path("", save_path, "config.yaml"))  # pyright: ignore
- 
+
     with TemporaryDirectory() as output_dir:
 
         training_args = TrainingArguments(
@@ -192,8 +203,8 @@ def run_train(config: TrainConfig):
         trainer = Trainer(
             model=model,
             args=training_args,
-            train_dataset=formatted_dataset["train"],
-            eval_dataset=formatted_dataset["validation"],  # pyright: ignore
+            train_dataset=train_dataset,
+            eval_dataset=valid_dataset, 
             tokenizer=processor.tokenizer,
             #Collator is not needed as we are doing batch size 1 for now...
             #data_collator=collator,
@@ -215,9 +226,8 @@ def run_train(config: TrainConfig):
                 logger.info("LoRA adapters merged successfully.")
 
             model.save_pretrained(best_dir)
-
-            logger.info("Saved best model to %s", best_dir)
 
+            logger.info("Saved best model to %s", best_dir)
 
         # Uncomment to test speed of data loader
         # train_dataloader = DataLoader(formatted_dataset["train"], batch_size=1, num_workers=4, shuffle=False)
@@ -232,4 +242,4 @@ def main():
 
 
 if __name__ == "__main__":
-    main()
+    main()

scripts/qwen2vl-7b-gantry.sh

@@ -19,7 +19,7 @@ run_name=$(basename "$0" .sh)
 # --cluster 'ai2/allennlp-cirrascale' \
 # --priority high \
 
-CLUSTER='jupiter'
+CLUSTER='pluto'
 
 gantry run \
     --description "${run_name}"\