ProofKV-scorer
a model for compiling proof chains into KV memory
2 min readupdated
ProofKV-scorer is a set of cross-encoder checkpoints built on microsoft/MiniLM-L12-H384-uncased with roughly 33 million parameters, created for the ProofKV research project from an EACL 2027 paper. The model reads a document once, finds and scores relevant text spans and the links between them, and builds a reusable evidence base for question-answering (QA) systems. Unlike standard full-context scanning methods, it selects chains of facts under a strict KV memory budget and passes them to a decoder. It is useful for researchers and developers working on multi-hop retrieval across long documents and LLM optimization.
Key features
- Three specialized heads: start and end heads for scoring text spans against a question, plus a nine-position linking head for paired spans.
- Compact size: a 12-layer architecture with 384 hidden states (~33M parameters), initialized from MiniLM.
- Ready-to-use checkpoints: a final interpolated version for broad task coverage and a separate checkpoint trained on MuSiQue replays.
- Repository integration: full compatibility with the ProofKV release scripts for selector verification and testing.
What you can do
- QA research: reproduce experiments from the paper on compiling proof chains for structured answer retrieval.
- Decoder integration: pair the scorer with models like Qwen2.5-14B to test packed contexts and distractor handling.
- Retrieval analysis: run missing-hop evaluations, iterative search, and fixed-byte-budget matrix generation.
Example
from huggingface_hub import snapshot_download
snapshot_download("XingjianZhangDev/ProofKV-scorer", local_dir="checkpoints",
allow_patterns=["proofkv_scorer_*/*"])
from proofkv.proofcache.model import TypedRelationScorer
scorer, tokenizer = TypedRelationScorer.load_checkpoint("checkpoints/proofkv_scorer_final", device="cuda")Read more
DetailsBenchmarks: five-task suite from the paper, including HotpotQA, 2WikiMultiHopQA, and MuSiQue.
Training: AdamW optimizer with a 2e-5 learning rate, 0.01 weight decay, batch size 48, and 1.0 gradient clipping.
Limitations: maximum input length is 320 tokens; the model is designed for English.
License: MIT.










Comments