Merge pull request #8 from shahules786/finetune-emb

Embedding finetuning experiments

Author: shahules786

.gitignore

@@ -3,7 +3,7 @@
 # python
 **/.ipynb_checkpoints
 .python-version
-
+finetune-embedding/wikidata
 # integrations
 **/.chroma/index
 **/wandb

finetune-embedding/data_creation.ipynb

@@ -0,0 +1,96 @@
+
+from ragas.metrics.critique import AspectCritique
+from llama_index import load_index_from_storage
+from llama_index import  StorageContext, set_global_service_context, ServiceContext
+from datasets import Dataset
+from langchain.embeddings import HuggingFaceEmbeddings
+from llama_index.storage.docstore import SimpleDocumentStore
+from llama_index.storage.index_store import SimpleIndexStore
+from llama_index.vector_stores import SimpleVectorStore
+import json
+from tqdm import tqdm
+import os
+
+def write_to_json(filename,finetuning_dataset):
+    
+    database = json.load(open(filename))
+    database.extend(finetuning_dataset)
+    with open(filename,'w') as file:
+        json.dump(database, file, indent=4)
+
+if __name__ == "__main__":
+    
+    embed_model = HuggingFaceEmbeddings(model_name="BAAI/bge-small-en-v1.5")
+    service_context = ServiceContext.from_defaults(embed_model=embed_model,)
+    critic = AspectCritique(name="filter", definition="Does the submission contain information that can be derived from input?")
+    dataset = Dataset.from_json("wikidata/indices/subset.json")
+    filename = "finetuning_dataset.json"
+    
+    if not os.path.exists(filename):
+        with open(filename,"w") as file:
+            json.dump([], file)
+        
+    
+    
+    batch_size=100
+    max_ragas_score = 0.8
+    threshold=0.8
+    for batch in tqdm(range(0,len(dataset)+1, batch_size)):
+        datapath=f"./sample-{batch}.index/"
+        # create storage context using default stores
+        storage_context = StorageContext.from_defaults(
+            docstore=SimpleDocumentStore.from_persist_dir(persist_dir=datapath),
+            vector_store=SimpleVectorStore.from_persist_dir(persist_dir=datapath),
+            index_store=SimpleIndexStore.from_persist_dir(persist_dir=datapath),
+        )
+        set_global_service_context(service_context)
+        index = load_index_from_storage(storage_context)
+        retriever = index.as_retriever(similarity_top_k=1)
+        subsample = dataset.select(range(batch, min(len(dataset), batch+batch_size)))
+        finetuning_dataset = []
+
+        try:
+            for item in subsample:
+                if item["ragas_score"] <= max_ragas_score:
+        
+                    node = retriever.retrieve(item["Answer"])[0]
+                    filter = critic.score_single({"question":node.get_content(),"answer":item["Answer"]})
+
+                    # if node.get_score()>=threshold:
+                    if filter:
+                        pos_chunk = node.to_dict()
+                    else:
+                        continue
+
+                    
+        
+                    retrieved_chunks = item["chunks"]
+                    # hard negatives : till positive hash
+                    hard = True
+                    hard_negatives,negatives = [], []
+                    for node in retrieved_chunks:
+        
+                        if node["node"]["hash"] == pos_chunk["node"]["hash"]:
+                            hard = False
+                            continue
+        
+                        if hard:
+                            hard_negatives.append(node)
+                        else:
+                            negatives.append(node)
+        
+                    sample = {"Question":item["Question"], "Answer":item["Answer"],
+                            "Context":item["Context"],
+                            "Conversation_no":item["Conversation_no"],
+                            "Turn_no":item["Turn_no"],
+                            "Positives":[pos_chunk["node"]["text"]],
+                            "Negatives":[chunk["node"]["text"] for chunk in negatives],
+                            "Hard_negatives":[chunk["node"]["text"] for chunk in hard_negatives]}
+                    finetuning_dataset.append(sample)
+                    
+            write_to_json(filename, finetuning_dataset)
+        except Exception as e:
+            print(e)
+                
+                
+