---
pretty_name: RPMem Training and Benchmark Data
language:
- en
tags:
- rpmem
- parametric-memory
- conversational-memory
configs:
- config_name: compiler_sessions
  data_files:
  - split: records
    path: compiler/sessions/*.parquet
- config_name: compiler_probes
  data_files:
  - split: records
    path: compiler/probes/*.parquet
- config_name: compiler_indices
  data_files:
  - split: train
    path: compiler/indices/train/*.parquet
  - split: query_validation
    path: compiler/indices/query_validation/*.parquet
  - split: validation
    path: compiler/indices/validation/*.parquet
---

# RPMem Training and Benchmark Data

Prepared data for **RPMem: Learning Recurrent Parametric Memory Across Sessions
for LLM Agents**. [Code and training instructions](https://github.com/Quark-Medical/rpmem).

## Contents

| Directory | Contents |
| --- | --- |
| `compiler/` | Sessions, probes, reference answers, and split indices (2.9 GB) |
| `benchmarks/perma/` | Prepared PERMA data |
| `benchmarks/personamem_v2/formal_v1/` | Prepared PersonaMem-v2 data |
| `benchmarks/prefeval/formal_v1/` | Prepared PrefEval data |

Total: **4.3 GB**. Teacher caches and model weights are not included.

## Download

```bash
hf download PolarSnowLeopard/RPMem-data --repo-type dataset --local-dir data/rpmem
```

For a partial download, add `--include "compiler/**"` or `--include "benchmarks/**"`.

## Usage

For [compiler training](https://github.com/Quark-Medical/rpmem/blob/main/docs/reproduction.md),
use `compiler/train.corpus.json` with the supplied split indices, not the Hub
viewer's `records` split. Teacher targets are generated using the code repository.

For [benchmark training and evaluation](https://github.com/Quark-Medical/rpmem/blob/main/docs/benchmarks.md),
use the directories above and skip raw-data preparation. To reproduce the
historical paper runs, select `--first-session-rule gate_zero_state`
(`FIRST_SESSION_RULE=gate_zero_state` for the PERMA launcher).

## Sources

See [SOURCES.md](SOURCES.md) for upstream datasets and their respective terms.