| # Where the input files come from | |
| `corpus.jsonl` and `corpus-dp1.jsonl` are replay corpora built from recorded coding-agent sessions in the [LMCache agentic traces](https://huggingface.co/datasets/sammshen/lmcache-agentic-traces) dataset. The five source parquet shards are byte-identical to that dataset's `data/train-0000{0..4}-of-00005.parquet` files; their SHA-256 hashes are listed in `corpus.manifest.json` and `corpus-dp1.manifest.json`. The shards themselves are not re-published here — download them from the source dataset. | |
| How the corpora were built from the traces: recorded assistant messages are dropped (the live model under test regenerates them during replay), tool-role messages become user-role turns wrapped as execution results, and the recorded gaps between calls become replay delays. Sessions that would break prompt-cache realism are skipped, never repaired. `corpus.jsonl` targets an 8-rank data-parallel server; `corpus-dp1.jsonl` is the same build for a single-rank server. Full build settings and skip counts are in each corpus manifest. | |
| `rhythm.json` is the arrival-time pattern the paced runs replay, extracted from the same recorded sessions. | |