Merge pull request #47 from arcee-ai/chore/training-experiment

Chore/training experiment

Author: Ben-Epstein

dalm/datasets/qa_gen/question_answer_generation.py

@@ -118,7 +118,8 @@ def generate_qa_from_dataset(
     # shuffle data
     dataset.shuffle(seed=42)
     # select a subset
-    small_dataset = dataset.select(range(sample_size))
+    num_samples = min(sample_size, len(dataset))
+    small_dataset = dataset.select(range(num_samples))
     # train-test split
     small_dataset_splits = split_dataset(small_dataset, title_column_name)
     print(

dalm/training/rag_e2e/train_rage2e.py

@@ -513,7 +513,8 @@ def train_e2e(
                 )
                 generator_tokenizer.save_pretrained(generator_ckpt_path)
             accelerator.wait_for_everyone()
-    accelerator.end_training()
+    if with_tracking:
+        accelerator.end_training()
 
 
 def main() -> None:

dalm/training/retriever_only/train_retriever_only.py

@@ -242,20 +242,20 @@ def train_retriever(
             passage_max_len=passage_max_len,
         ),
         batched=True,
-        remove_columns=dataset["train"].column_names,
+        remove_columns=dataset.column_names,
         desc="Running tokenizer on dataset",
     )
 
     # Log a few random samples from the training set:
-    for index in random.sample(range(len(processed_datasets["train"])), 3):
-        logger.info(f"Sample {index} of the training set: {processed_datasets['train'][index]}.")
+    for index in random.sample(range(len(processed_datasets)), 3):
+        logger.info(f"Sample {index} of the training set: {processed_datasets[index]}.")
 
     model.print_trainable_parameters()  # type: ignore # No idea what mypy is complaining about.
     accelerator.print(model)
 
     # get dataloaders
     train_dataloader = DataLoader(
-        processed_datasets["train"],
+        processed_datasets,
         shuffle=True,
         collate_fn=default_data_collator,
         batch_size=per_device_train_batch_size,
@@ -308,7 +308,7 @@ def train_retriever(
         accelerator.register_load_state_pre_hook(load_model_hook)
 
     logger.info("***** Running training *****")
-    logger.info(f"  Num examples = {len(processed_datasets['train'])}")
+    logger.info(f"  Num examples = {len(processed_datasets)}")
     logger.info(f"  Num Epochs = {num_train_epochs}")
     logger.info(f"  Instantaneous batch size per device = {per_device_train_batch_size}")
     logger.info(f"  Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}")
@@ -409,7 +409,8 @@ def train_retriever(
                 )
                 tokenizer.save_pretrained(output_dir)
             accelerator.wait_for_everyone()
-    accelerator.end_training()
+    if with_tracking:
+        accelerator.end_training()
 
 
 def main() -> None:

experiments/README.md

@@ -0,0 +1,91 @@
+# Testing e2e retriever hit rate
+
+
+Inspired by [llama-index](https://gpt-index.readthedocs.io/en/latest/examples/finetuning/embeddings/finetune_embedding_adapter.html)
+
+# Setup
+
+```shell
+pip install indomain
+pip install -r requirements.txt
+python data_gen.py
+mkdir qa-outputs
+```
+
+## Create the datasets
+First, create the train dataset
+```shell
+dalm qa-gen train_data.csv --output-dir qa-outputs --passage-column-name text --title-column-name title --sample-size 1000000
+```
+This creates a train and test file (because we typically want to split), so merge those into 1
+```shell
+head -n 1 qa-outputs/question_answer_pairs_train.csv > question_answer_pairs.csv && tail -n+2 -q qa-outputs/*.csv >> question_answer_pairs.csv
+rm qa-outputs/*.csv
+mv question_answer_pairs.csv qa-outputs
+```
+
+Same for the validation data
+```shell
+dalm qa-gen val_data.csv --output-dir qa-outputs-test --passage-column-name text --title-column-name title --sample-size 100000
+head -n 1 qa-outputs-test/question_answer_pairs_train.csv > question_answer_pairs_test.csv && tail -n+2 -q qa-outputs-test/*.csv >> question_answer_pairs_test.csv
+rm -rf qa-outputs-test
+mv question_answer_pairs_test.csv qa-outputs
+```
+
+Now we have 2 files for training and eval
+```shell
+(.venv) root@f4ec1ae23983:# ls -lash qa-outputs/
+total 2.3M
+1.4M -rw-r--r-- 1 root root 1.4M Sep 20 20:02 question_answer_pairs.csv
+956K -rw-r--r-- 1 root root 953K Sep 20 20:14 question_answer_pairs_test.csv
+```
+
+## Rage2e training
+
+Then we train e2e
+```shell
+dalm train-rag-e2e \
+"qa-outputs/question_answer_pairs.csv" \
+"BAAI/bge-small-en" \
+"meta-llama/Llama-2-7b-hf" \
+--dataset-passage-col-name text \
+--output-dir "rag_e2e_checkpoints_bgsmall" \
+--no-with-tracking \
+--per-device-train-batch-size 12
+```
+
+And eval
+```
+python ../dalm/eval/eval_retriever_only.py  --dataset_path qa-outputs/question_answer_pairs_test.csv --retriever_model_name_or_path "BAAI/bge-small-en" --passage_column_name text --query_column_name Question --retriever_peft_model_path rag_e2e_checkpoints_bgsmall/retriever --embed_dim 384
+
+*************
+Retriever results:
+Recall: 0.8202054794520548
+Hit Rate: 0.8202054794520548
+*************
+```
+
+
+
+## Retriever only training
+
+Train the retriever only
+```
+dalm train-retriever-only "BAAI/bge-small-en" "qa-outputs/question_answer_pairs.csv" \
+--output-dir "retriever_only_checkpoints_bgsmall" \
+--use-peft \
+--dataset-passage-col-name text \
+--per-device-train-batch-size 150
+```
+
+and eval
+```
+python ../dalm/eval/eval_retriever_only.py  --dataset_path qa-outputs/question_answer_pairs_test.csv --retriever_model_name_or_path "BAAI/bge-small-en" --passage_column_name text --query_column_name Question --retriever_peft_model_path retriever_only_checkpoints_bgsmall/ --embed_dim 384
+
+*************
+Retriever results:
+Recall: 0.8116438356164384
+Precision: 0.08116438356164453
+Hit Rate: 0.8116438356164384
+*************
+```

experiments/data_gen.py

@@ -0,0 +1,39 @@
+import json
+
+from llama_index import SimpleDirectoryReader
+from llama_index.node_parser import SimpleNodeParser
+from llama_index.schema import MetadataMode
+import pandas as pd
+
+TRAIN_FILES = ["uber_2021.pdf"]
+VAL_FILES = ["lyft_2021.pdf"]
+
+
+def load_corpus(files, verbose=False):
+    if verbose:
+        print(f"Loading files {files}")
+
+    reader = SimpleDirectoryReader(input_files=files)
+    docs = reader.load_data()
+    if verbose:
+        print(f"Loaded {len(docs)} docs")
+
+    parser = SimpleNodeParser.from_defaults(chunk_size=512)
+    nodes = parser.get_nodes_from_documents(docs, show_progress=verbose)
+
+    if verbose:
+        print(f"Parsed {len(nodes)} nodes")
+
+    return nodes
+
+
+
+train_nodes = load_corpus(TRAIN_FILES, verbose=True)
+val_nodes = load_corpus(VAL_FILES, verbose=True)
+
+train_df = pd.DataFrame({"text": [node.text for node in train_nodes], "title": [node.id_ for node in train_nodes]})
+val_df = pd.DataFrame({"text": [node.text for node in val_nodes], "title": [node.id_ for node in val_nodes]})
+
+train_df.to_csv("train_data.csv")
+val_df.to_csv("val_data.csv")
+

experiments/lyft_2021.pdf

@@ -0,0 +1,2 @@
+llama-index
+pypdf