# Sources and Attribution

The supplied release preserves source metadata and paths without claiming new
ownership over upstream content. Source descriptions are separated from licenses.

| Source | Role in Beyond Recall | Authoritative project |
|---|---|---|
| LoCoMo | Carrier conversations used as long-history context, as described in the paper | [snap-research/locomo](https://github.com/snap-research/locomo) |
| MMDU | Adapted multimodal carrier conversations, as described in the paper | [Liuziyu77/MMDU](https://github.com/Liuziyu77/MMDU) |
| MMLongBench-Doc | Source document annotations used to select PDF evidence pages | [mayubo2333/MMLongBench-Doc](https://github.com/mayubo2333/MMLongBench-Doc) |
| Collected task media and human reconstruction | Media-grounded task design, delayed queries and reference answers | Accompanying Beyond Recall manuscript |

LoCoMo authors: Adyasha Maharana, Dong-Ho Lee, Sergey Tulyakov, Mohit Bansal,
Francesco Barbieri and Yuwei Fang. Paper: Evaluating Very Long-Term Conversational
Memory of LLM Agents (ACL 2024). See its official repository for the reference.

MMDU paper: [MMDU: A Multi-Turn Multi-Image Dialog Understanding Benchmark and
Instruction-Tuning Dataset for LVLMs](https://arxiv.org/abs/2406.11833).

MMLongBench-Doc paper: [MMLongBench-Doc: Benchmarking Long-context Document
Understanding with Visualizations](https://arxiv.org/abs/2407.01523).

PDF modifications retain source-annotated evidence, boundary pages and optional
other pages. `benchmark/PDF_PREPROCESSING.json` records retained page numbers and
checksums. Dialogue speaker identifiers, task paths and question/answer records
preserve the release schema. Packaged videos retain their original audio
tracks.

## Public media inventory

`metadata/MEDIA_MANIFEST.csv` lists all 4,809 packaged media paths, media types
and SHA256 hashes. The 93 PDF paths have confirmed source-dataset and document
matches. Other 4,716 paths explicitly use `unknown` for unconfirmed sources.
Repeated paths remain separate entries even when their file bytes are identical.

The manifest records known source identification, not a complete original-URL
catalog or a redistribution-rights certification. A source match does not itself
grant reuse rights. Source-specific conditions remain described in
[Licensing](LICENSING.md); unknown attribution must not be interpreted as
public-domain status.

The 93 PDF source matches are also recorded in `metadata/PDF_SOURCE_MAP.json`,
with original-source and current cropped-file hashes distinguished.
`benchmark/PDF_PREPROCESSING.json` retains original page mappings and cropping
information. These version and provenance records are distinct from the media
inventory's known/unknown source classification.