LafouCC/longlamp-ag-stream

LongLaMP abstract-generation — per-user update streams

A re-curation of LongLaMP/LongLaMP, config abstract_generation_temporal into per-user update streams: each author's whole publication history, in chronological order, packed into disjoint chunks, where every item is both a condition item and a possible target. Built for training a recurrence across adapter updates, θ_t = f(chunk_t, θ_{t-1}), rather than a single condition→output pass.

Splits are by author, 22,421 train / 200 valid / 200 test.

The target prompt is LongLaMP's own, key items included:

Generate an abstract for the title "{title}" using the following items: {clause}

The benchmark ships that clause only for its ~29 k designated targets, and a stream scores history papers, so the clause was regenerated for all 1,701,966 corpus papers with the paper's verbatim extraction prompt. 99.8 % of items carry one; the rest fall back to Generate an abstract for a paper with the title: {title}, per row. See Key items below and KEY_ITEMS_HANDOFF.md.

Files

file	rows	what
streams/stream_{train,valid,test}_v3.parquet	29,641 / 200 / 200	the dataset. One row per (author, episode)
key_items-00{0..5}-of-006.parquet	283,661 each	paper_id → keywords, key_items, ok, raw. The regenerated clause for all 1,701,966 papers
authors_v2.parquet	22,821	one row per author: chronological items with topic, seen_in_train and key_items filled. The curator's direct input
papers-00{0..5}-of-006.parquet	283,661 each	1,701,966 distinct papers: paper_id, year, n_owners, title, abstract. Deduped across co-authors; the extraction input
key_items_gold.parquet	10,783	LongLaMP's real clause for papers that are also in this corpus. Validation only — 0.5 % of the corpus, never a data source
key_items_gold_pred.parquet	10,783	ours for those same papers, for the side-by-side
key_items_gold_eval.json	—	the fidelity table below
key_items.verify.json	—	per-shard id-set check against papers-*
topics.parquet	1,701,966	paper_id → topic, k-means over title embeddings (k=256)
train_paper_ids.parquet	1,683,875	papers owned by a train author; source of seen_in_train
split_v3.json	22,821	user_id → train|valid|test, plus cohort sizes

How it was built

1. Ingest. The benchmark's own splits are a temporal cut over the same people — val and test author sets are strict subsets of train's — so they are not a user partition and cannot be used as one. Pooled and re-split by author. For all 22,821 authors the largest profile equals the union of their profiles across splits, so one row per author is the complete history. Profiles are newest-first in the source; reversed here.
2. Topics. Papers have no category, so a recurring attribute is derived: k-means (k=256) over title embeddings (Qwen3-Embedding-0.6B). Titles rather than abstracts, so the label is a function of the prompt side, not the gold. Largest cluster is 1.24 % of papers.
3. Key items. See below.
4. Split. Stable SHA-512 hash of (tag, seed, user_id), sliced. Fixed cohorts of 200 valid / 200 test, matching P2P's own test-set sizes; everything else is train.
5. Chunking. Greedy oldest-first packing under min(8 items, 4096 tokens) measured through the condition rendering. An episode is a window over the chunk sequence; state t reads chunks[0..t-1] and is scored on every item of chunks[t]. Chunks within an episode are disjoint — otherwise the state is partly re-derivable from the current chunk and a stateless baseline gets a free pass. Train windows cap the unroll at T_max = 16; valid/test unroll the author's whole stream as one episode.

Key items

LongLaMP's five keywords are extracted from the paper's own abstract — the gold output — by an LLM, with this prompt (arXiv 2407.11016, verbatim):

Mention 5 short keywords of the following abstract of the paper that shows their main findings and claims: [Abstract] [OUTPUT] Keyword 1, Keyword 2 , Keyword 3 , Keyword 4, Keyword 5

That content leak is the benchmark's design and it is reproduced on purpose. Making the keywords title-derived instead would be a different, harder task. Two things follow, and both belong in any number reported off this corpus:

• Absolute ROUGE is inflated relative to a title-only variant — the task is partly extractive. Do not compare across the two.
• The leak is uniform across arms, so a between-arm delta is still clean.

The paper never names the extraction model. Their repo's only LLM is gpt-35-turbo via Azure, so GPT-3.5-turbo is the strong inference and nothing more. This used Qwen2.5-7B-Instruct, bf16, greedy, max_tokens=48, seed 42 under vLLM 0.29.0, abstracts truncated to 512 tokens (0.33 % hit the cap), shape=comma as the paper specifies. The substitution is the main threat to fidelity, so it is checked against the 10,783 papers carrying LongLaMP's real clause:

check	LongLaMP (gold)	this corpus
parse rate	—	0.9985 (target ≥ 0.95)
keyword words mean / p50 / p90	2.31 / 2 / 4	2.35 / 2 / 4
5 items per paper	96.7 %	90.5 %
token-F1 vs gold clause	—	0.52
ROUGE-1 recall	—	0.65
gold keywords verbatim	—	0.43

The overlap row is a sanity band, not a target — two different LLMs asked for "5 short keywords" agree loosely, and optimising it would be overfitting to one unknown model's idiosyncrasies. The length row is the one that matters: a generator emitting sentence-long "keywords" would leak more of the abstract into the prompt than LongLaMP leaked, changing the task's difficulty.

Corpus-wide: 1,701,966 papers, ok rate 0.9979, keyword length mean 2.29 / p50 2 / p90 3. Every shard's id set was checked against its papers-* source — 0 duplicates, 0 missing.

Key items are TARGET-SIDE ONLY. The condition rendering (Paper Title: … | Abstract: …) does not carry them, and the packer measures its token budget through that rendering, so regenerating or re-joining key items moves no chunk boundary. The raw column in key_items-* keeps each model output so the clause can be re-parsed without re-running the extraction.

Numbers

	train	valid	test
users	22,421	200	200
episodes	29,641	200	200
states	318,963	2,654	2,733
candidate targets	2,470,649	20,509	21,104
mean T	10.76	13.27	13.66
chunks hitting the token cap	0.0 %	0.0 %	0.0 %
states with a retention target	0.698	0.721	0.732
items carrying key items	99.81 %	99.79 %	99.84 %

The 4096-token cap never binds: an 8-item chunk of title+abstract averages ~1,620 tokens.

Three things to state whenever you report a number

1. Co-authorship is a memorisation channel, flagged not filtered. Co-authors are both users here, so 39.1 % of the 1,701,966 distinct papers have more than one owner, and the co-authorship graph is a single component covering 99.7 % of authors — a leak-free author partition does not exist. Measured on these files, a held-out author's papers are train-visible for 57.7 % (valid) / 61.0 % (test) of items; per author, mean 54.0 % / 57.8 % (p10 0.19 / p90 0.85). That tracks n_owners > 1 almost exactly (0.582 / 0.615), which is the mechanism: a held-out paper is train-visible essentially iff it has a co-owner in the corpus, and train is 98.2 % of authors.

Do not quote the corpus-wide seen_in_train (0.993) as the leak rate. It is dominated by train rows, where the flag is vacuously true. Only the valid/test slice means anything. An earlier version of this card reported ~99 % for this reason; the numbers above are the corrected ones.

Every item carries n_owners and seen_in_train so the overlap stays a slice the read-out can cut. This is inherited from the benchmark, not introduced here: P2P's own split has it too — 200/200 of its random_test / ood_test users hold papers that appear in a train user's history. Neither P2P nor LongLaMP reports a seen/unseen slice.

2. The target prompt carries key items drawn from the gold abstract. See Key items above. A number here is not comparable to a published LongLaMP number (their pipeline differs), and it is not comparable to the title-only variant of this same corpus.

3. Time is year-granular and timestamp is synthetic. The source carries year and nothing finer; items of one (author, year) are spread evenly across that year in source order. The year boundary is exact, everything below it is a placeholder.

Schema — stream rows

user_id          string
n_shared_items   int32     items with n_owners > 1
n_items          int32
ep_idx           int32
start_chunk      int32     position in the author's own chunk sequence
n_chunks_user    int32
T                int32     supervised states; state t reads chunks[0..t-1]
chunks           list<list<item>>
chunk_meta       list<{ts_start, ts_end, n_items, n_tokens, token_capped}>
ret_idx          list<int32>          per state: index in chunks[t] of the
ret_dist         list<int32>          reserved retention target, or -1
ret_all          list<list<int32>>    per state, per target: retention distance
                                      d = (t-1) - j; 0 = revision, >=1 = retention,
                                      -1 = novel
item = {timestamp, year, paper_id, topic, n_owners, seen_in_train,
        title, abstract, key_items}

abstract is stored untruncated. It is capped at 800 characters when rendered as a condition and left whole as gold — do not apply the cap to the target. key_items is the clause that follows using the following items: ; empty means fall back to the title-only prompt for that row.

The parquet's schema metadata carries corpus, task, item_format and prompt_variant. That is the receipt: the renderer is not recoverable from the data, and using the wrong one silently produces empty text or the wrong prompt. These files are prompt_variant=key_items.

Source and license

Derived from LongLaMP/LongLaMP (abstract_generation_temporal), itself built from the Citation Network Dataset (V14) filtered to authors with ≥ 70 publications. See LongLaMP: A Benchmark for Personalized Long-form Text Generation.

@misc{kumar2024longlampbenchmarkpersonalizedlongform,
  title={LongLaMP: A Benchmark for Personalized Long-form Text Generation},
  author={Ishita Kumar and Snigdha Viswanathan and Sushrita Yerra and Alireza Salemi
          and Ryan A. Rossi and Franck Dernoncourt and Hanieh Deilamsalehy and
          Xiang Chen and Ruiyi Zhang and Shubham Agarwal and Nedim Lipka and
          Chien Van Nguyen and Thien Huu Nguyen and Hamed Zamani},
  year={2024}, eprint={2407.11016}, archivePrefix={arXiv}, primaryClass={cs.CL}
}