xbridge-eval-cache / README.md
wyangw's picture
Upload folder using huggingface_hub
6c006eb verified
|
Raw
History Blame Contribute Delete
1.77 kB
metadata
license: apache-2.0
task_categories:
  - question-answering
tags:
  - xbridge
  - llama
  - eval-cache

XBridge Evaluation Caches (Llama-3.1-8B sender)

Precomputed evaluation caches for the XBridge repo. Each file stores, for a fixed set of eval items per task, the sender's (Llama-3.1-8B-Instruct) last-layer hidden states H_S and attention-sorted context token IDs — the inputs the Latent Enrichment Bridge (LEB) module queries at inference time. This is a cache of a Llama forward pass, not a copy of the underlying QA datasets themselves.

File Task Source dataset Eval items
countries_task_split.pt countries local (dataloader/data/countries.jsonl) 100
tipsheets_task_split.pt tipsheets local (dataloader/data/tipsheets.jsonl) 200
twowikimqa_task_split.pt twowikimqa Xnhyacinth/LongBench (2wikimqa, test) 100
multifieldqa_en_task_split.pt multifieldqa_en Xnhyacinth/LongBench (multifieldqa_en, test) 75
hotpotqa_attnselect.pt hotpotqa hotpot_qa (distractor, train, seed=42 shuffle, first 200) 200

Usage

Download into precomputed/ at the repo root:

huggingface-cli download wyangw/xbridge-eval-cache --repo-type dataset --local-dir precomputed

Then run Option A from the repo README.

Regenerating from scratch

Each file can also be rebuilt from public sources using the scripts in the repo (scripts/precompute_task_split.py, scripts/precompute_tipsheets_split.py, scripts/precompute_hotpotqa_eval.py) — requires meta-llama/Llama-3.1-8B-Instruct access (gated on HF) and a GPU. See the repo README for exact commands.