hit@5 float64 | recall@5 float64 | questions int64 | seconds float64 | ljos string | ljos_commit string | packset string | dataset_repo string | dataset_revision string | cpu_model string | date timestamp[s] |
|---|---|---|---|---|---|---|---|---|---|---|
0.959574 | 0.909113 | 470 | 75.095 | 0.41.11 | 62965c1070d003126d959dc25399be64cec3128b | 0.17.6 | xiaowu0162/longmemeval-cleaned | 98d7416c24c778c2fee6e6f3006e7a073259d48f | Intel(R) Xeon(R) Processor | 2026-11-13T00:00:00 |
ljos LongMemEval_S retrieval
This dataset is one BM25+ session-retrieval run of the ljos scorer on cleaned LongMemEval_S.
The scorer is longmemeval_s_recall in the ljos tree.
It calls packset-core's BM25+ on the CPU.
It loads no encoder, downloads no model, and asks no remote judge.
Hit@5 is the share of scored questions with at least one evidence session among the first five ranks.
Recall@5 is the mean, over those questions, of the share of evidence sessions found in the top five.
Abstentions, whose ids end in _abs, are left out.
The questions are not in this dataset.
They stay in xiaowu0162/longmemeval-cleaned at revision 98d7416c24c778c2fee6e6f3006e7a073259d48f, file longmemeval_s_cleaned.json.
That dataset's card says license: mit, so a copy is allowed.
This dataset does not copy the file.
The run names that revision and stops unless the sha256 is d6f21ea9d60a0d56f34a05b609c79c88a451d2ae03597821ea3d5a9678c3a442.
Files
out/summary.json records the BM25+ figures for this run.
hit@5 and recall@5 are the six-decimal ratios the scorer prints.
questions is the number of scored questions.
seconds is the scorer's wall time, three digits past the point, from after the file is on disk through the scores.
ljos is the ljos version in the checkout.
ljos_commit is that checkout's git commit.
packset is the packset-core version the scorer links.
dataset_repo and dataset_revision name the upstream file.
cpu_model is the model name line from /proc/cpuinfo, or unknown when that file has none.
date is the day the run finished, in Coordinated Universal Time.
out/per_question.csv is one row per scored question, BM25+ only.
question_id and question_type identify the row.
evidence is how many labelled sessions the question has.
found@1, found@5 and found@10 count how many of those sessions sit inside that depth.
hit@1, hit@5 and hit@10 are 1 when the matching found column is above zero.
recall@1, recall@5 and recall@10 are found divided by evidence, six digits past the point.
The summary hit@5 equals the fraction of rows with hit@5 of 1.
The summary's recall@5 is the mean of the unrounded per-question shares, then printed at six digits.
Averaging the already-printed recall@5 column can differ in the last digit.
run.sh is the command that fetches the pinned file and runs the scorer.
README.md is this card.
Reproduce
From a checkout of https://github.com/leidarljos/ljos:
sh bench/hf/run.sh
The command needs the Rust toolchain named in rust-toolchain.toml, curl, and sha256sum or shasum.
It writes bench/hf/out/summary.json and bench/hf/out/per_question.csv.
Licence
The Massachusetts Institute of Technology licence.
Copyright © 2026, leiðarljós Developers.
The ljos sources are at https://github.com/leidarljos/ljos.
- Downloads last month
- 52