JJTsao/fine-tuned_movie_retriever-all-minilm-l6-v2

🎬 Fine-Tuned Movie Retriever (Rich Semantic & Metadata Queries + Smart Negatives)

This is a custom fine-tuned sentence-transformer model designed for movie and TV recommendation systems. Optimized for high-quality vector retrieval in a movie and TV show recommendation RAG pipeline. Fine-tuning was done using ~32K synthetic natural language queries across metadata and vibe-based prompts:

• Enriched vibe-style natural language queries (e.g., Emotionally powerful space exploration film with themes of love and sacrifice.)
• Metadata-based natural language queries (e.g., Any crime movies from the 1990s directed by Quentin Tarantino about heist?)
• Smarter negative sampling (genre contrast, theme mismatch, star-topic confusion)
• A dataset of over 32,000 triplets (query, positive doc, negative doc)

🧠 Training Details

• Base model: sentence-transformers/all-MiniLM-L6-v2
• Loss function: MultipleNegativesRankingLoss
• Epochs: 4
• Optimized for: top-k semantic retrieval in RAG systems

📈 Evaluation: Fine-tuned vs Base Model

Metric	Fine-Tuned Model Score	Base Model Score
Recall@1	0.428	0.149
Recall@3	0.657	0.258
Recall@5	0.720	0.309
Recall@10	0.795	0.382
MRR	0.563	0.230

Evaluation setup:

• Dataset: 3,600 held-out metadata and vibe-style natural queries
• Method: Top-k ranking using cosine similarity between query and positive documents
• Goal: Assess top-k retrieval quality in recommendation-like settings

📦 Usage

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("jjtsao/fine-tuned_movie_retriever")
query_embedding = model.encode("mind-bending sci-fi thrillers from the 2000s about identity")

🔍 Ideal Use Cases

• RAG-style movie recommendation apps
• Semantic filtering of large movie catalogs
• Query-document reranking pipelines

📜 License

Apache 2.0 — open for personal and commercial use.