vireshk/fine-tune-vs-rag-parity-index
Fine-Tune vs RAG — parity retrieval index The exact FAISS index the Fine-Tune vs RAG benchmark used for its rag-parity arm, published so the live demo retrieves over the same passages the report did. Not for clinical use. Not medical advice. Not a medical device. This is exam-explanation text from a public benchmark dataset, chunked for retrieval research. It contains the textual noise and errors documented in the report. What is in it File Contents… See the full description on the dataset page: https://huggingface.co/datasets/vireshk/fine-tune-vs-rag-parity-index.
Fine-Tune vs RAG — parity retrieval index
The exact FAISS index the Fine-Tune vs RAG benchmark used for its rag-parity arm, published so the live demo retrieves over the same passages the report did.
Not for clinical use. Not medical advice. Not a medical device. This is exam-explanation text from a public benchmark dataset, chunked for retrieval research. It contains the textual noise and errors documented in the report.
What is in it
- Source: the
exp(explanation) field of the MedMCQA training split only — 139,856 documents. No text derived from the benchmark's frozen test items is present; the repository's leakage test proves this against the committed split ids. - Chunking: as in
fvr.retrieval.corpus, mean 403 characters per chunk. - Embedder:
BAAI/bge-large-en-v1.5at revisiond4aa6901d3a41ba39fb536a557fa166f842b0e09, float16, normalised, with the document side embedded without the query instruction. Queries must be prefixed withRepresent this sentence for searching relevant passages:.
The name "parity" is the benchmark's: this corpus holds the same explanations the fine-tuned arm was trained on, so retrieval and fine-tuning are compared on identical information.
Loading
import faiss, json
from huggingface_hub import hf_hub_download
repo = "vireshk/fine-tune-vs-rag-parity-index"
index = faiss.read_index(hf_hub_download(repo, "index.faiss", repo_type="dataset"))
with open(hf_hub_download(repo, "passages.jsonl", repo_type="dataset")) as f:
passages = [json.loads(line) for line in f]Licence
MedMCQA is released under Apache 2.0 (Pal, Umapathi & Sankarasubbu, 2022); these chunks are a derived work under the same licence.
