annulus
Datasets
All datasets matching “annulus”annulus-ift-2000
Annulus Instruction-Tuning Data — Backbone (knowledge boundary = 2000)
Instruction fine-tuning data for the Annulus time-aware model's backbone (knowledge cutoff = year 2000). It teaches the model to (1) emit a leading control/year token and (2) produce the right behavior — recall a known fact, refuse or forecast beyond the boundary, or follow a year-agnostic instruction.
The cutoff is intrinsic to the model, NOT taught by the text. No sample states its knowledge boundary in… See the full description on the dataset page: https://huggingface.co/datasets/lfqian/annulus-ift-2000.annulus-eval
Annulus eval sweep results
Per-ckpt 3-task eval (PPL pre-2000 year-swap / ECON full 1980–2026 recall+prediction / CAP ICL)
for annulus runs v7, v8, v11 … Produced on misha (Yale) via econ_work/ pipeline.
results/sweep/ppl_to2005_*.txt — PPL per year-bucket (plain/correct/wnear/wfar)
results/econ_ab/ab_*_{recall,prediction}.{json,csv} — P(E) + A–E logits per tau (1980–2026)
results/captest_*_{plain,task1}.csv — cap ICL per-task (task1=带[Y2000], plain=不带)
dashboard.html — 3-tab… See the full description on the dataset page: https://huggingface.co/datasets/lfqian/annulus-eval.annulus-heldout
annulus-heldout
Held-out eval sets for the year-plugin project (same-source as 0X_filtered training, held out = not in training).
FY2001/heldout_2001_samesource.jsonl - 783 docs, all date=2001, same-source mix (common_corpus/USPTO/sec_edgar/arxiv). Use for FY2001 self-year absorption.
multiyear/eval_heldout.jsonl - ~10k docs, year/bucket_id/year_token, buckets 1500s..2000. Use for cross-year (far-year) cutoff.
Convention for later years
Add each new year's… See the full description on the dataset page: https://huggingface.co/datasets/lfqian/annulus-heldout.