LarrisX/packed-complaints

Packed complaints

A long context test for decision models: ask about every item in a long document, not one needle.

1,980 consumer complaints from the public CFPB database, each with one question: which kind of financial product is complaint N about? (nine options). The same complaints are asked alone, then packed into documents of 8k to 256k tokens, holding about 30 to 990 complaints each. Because every complaint is also asked alone, accuracy at each length pairs with the same model's score on the complaint by itself, so the gap measures only how well it finds and reads the right part of a long document.

A companion test packs long contracts instead of short complaints: LarrisX/packed-contracts.

Reference scores

Accuracy, one question per complaint, scored with Kev's own scorer (chance is 0.11). Complaints per input in brackets.

Model	Alone	32k (~123)	64k (~246)	128k (~490)	256k (~990)
Kev-0.8B 1.0	0.803	0.143	0.121	0.118	0.112
Kev-4B 1.0	0.841	0.491	0.420	—	—
Kev-9B 1.0	0.849	0.583	0.513	—	—
Kev-27B 1.0	0.857	0.853	0.854	—	—
Self-taught Kev-0.8B	0.804	0.798	0.791	0.788	0.748

From packed-complaints-v1 (its 32k and 64k documents are v0's). The bigger Kevs were scored to 64k.

Contents

Each folder is a suite in Kev's format (manifest.json + development.jsonl). Which one to score:

• Up to 64k: packed-complaints-v0-split with kev.benchmark --suite. The unsplit v0 holds up to ~250 questions per 64k document, more than Kev's scorer fits in GPU memory at once.
• Past 64k: packed-complaints-v1 with the repo's chunked scorer, which reads each document once and runs its questions in batches. packed-complaints-v1-check checks that scorer against Kev's own.

Folder	Lengths	Records	Questions	Notes
packed-complaints-v0	alone, 8k, 16k, 32k, 64k	2,101	9,817	the test; index.json maps each record to its length and complaints
packed-complaints-v0-split	same	2,303	9,817	each long document's questions in groups that share it, so Kev's scorer fits in GPU memory; answers equal the unsplit suite's
packed-complaints-v0-split-long	alone, 32k, 64k	2,190	5,912	the faster subset (8k and 16k run on Kev's slow exact path)
packed-complaints-v1	alone, 32k, 64k, 128k, 256k	2,016	13,806	past 64k: 8 documents of ~490 complaints and 4 of ~990, each length in two random orders
packed-complaints-v1-check	64k	2	516	two 64k documents with reference rows, to check a long-context scorer against Kev's

A record is one document (state) and its questions, each naming its complaint by ID:

{"state": "Complaint 4950009:\nHi. I would like to lodge a complaint against Coinbase ...\n\nComplaint 3370055:\n...",
 "questions": {"c4950009": {"type": "choice",
                            "instructions": "Which kind of financial product is complaint 4950009 about?",
                            "criteria": {"credit_reporting": "Credit reports, credit scores, or credit repair", "...": "..."},
                            "label": "money_transfer"}, "...": "..."}}

How it was built

• Complaints. CFPB narratives, 80–600 tokens, from shard 10 of davidheineman/consumer-finance-complaints-large @ 44cfa17, the snapshot Kev's documents-v1 was built from. Deduplicated against Kev's documents-v1 train set, balanced at 220 per product.
• Question and labels. Kev's documents-v1 product question, verbatim with its nine options. Labels are the consumer's own product choice, mapped to the options; three ambiguous products are skipped. No model is involved in labelling.
• Packing. Documents fill 80–93% of their nominal length under the Qwen3.5 tokenizer (Kev's longdoc-v1 rule), with a fresh random order at every length so neighbours change.

Licence

Apache-2.0 for the suite. The complaint narratives are US government work (public domain), and the question wording is Kev's (Apache-2.0).