emb-rag-minilm-relsft

LoRA adapter for MiniLM-L6-v2 — Relevance-SFT (direct, GISTEmbed tau0.02 lr1e-4).

Trained on multi-hop MuSiQue; evaluated on MuSiQue (in-domain) + 2WikiMultiHopQA, HotpotQA, NQ, TriviaQA (OOD).

Indirect feedback = the reward is a frozen reader LLM's answer success (log p(gold answer | query, retrieved context)) — no passage-level relevance labels. Rel-SFT = the direct-supervision baseline (contrastive InfoNCE on gold passages).

Full results (retrieval R@k + downstream RAG accuracy) and training scripts: https://irisicy4.github.io/exp-record/projects/indirect-rag/

from peft import PeftModel
from transformers import AutoModel, AutoTokenizer
base = AutoModel.from_pretrained("MiniLM-L6-v2")
model = PeftModel.from_pretrained(base, "Icey444/emb-rag-minilm-relsft").merge_and_unload()
tok = AutoTokenizer.from_pretrained("Icey444/emb-rag-minilm-relsft")