Blame for README.md

---
pretty_name: AgentLoad synthetic agent traces (PrefixAnvil)
license: other
task_categories: [other]
tags: [llm-inference, prefix-caching, kv-cache, agents, synthetic]
size_categories: [n<1K]
---

# AgentLoad synthetic agent traces

Synthetic OpenAI-compatible chat-completion request logs of tool-using agent sessions, with **ground-truth
labels of injected prefix-cache fragmentation** (F1–F8). Generated by `prefixanvil gen` (PrefixAnvil AgentLoad);
fully reproducible from a small JSON config and a seed. No real user data. The traces are generated by
PrefixAnvil — https://github.com/NagaYu/prefixanvil

## License

This dataset is made publicly visible solely so it can be read and evaluated. It is All Rights Reserved — no
license is granted beyond viewing and downloading for evaluation. Licensing or transfer inquiries:
https://nagayu-portfolio.nagayu.workers.dev/en

## Files

| file | content |
|---|---|
| `bench_main.jsonl` | 377 requests, 40 sessions, all eight fragmentation types injected at 6% each |
| `bench_main.labels.jsonl` | one line per injected event: `request_index`, `request_id`, `fault` (F1–F8), `scope` (within/cross session), `reference_request_id`, `location` |
| `bench_xsession.jsonl` | 155 requests, 30 sessions, no faults; each session's client lists the same tools in its own order |
| `configs/*.json` | the exact generator configs (seeds included) |

Each request line: `request_id`, `session_id`, `timestamp`, `model`, `messages`, `tools`.

## How it was generated

```bash
prefixanvil gen --config fixtures/bench/bench_main.json --out bench_main.jsonl
prefixanvil gen --config fixtures/bench/bench_xsession.json --out bench_xsession.jsonl
```

Byte-identical output for the same config (pure-Python RNG; checked in CI). File sha256 of the four `.jsonl`
files in this folder:

| file | sha256 |
|---|---|
| `bench_main.jsonl` | `cee5445df3364b3fc0a2af4a1f906356d537076857a5c4059ce96a6b800ef000` |
| `bench_main.labels.jsonl` | `5b114c0cf4547eadb77454313d82249e1fdccd869118d72864a31c16415b623a` |
| `bench_xsession.jsonl` | `271a8a6f3c2a61e9ae4b0878285e65ac96e17bf0b08202b415bff2c3a2f819df` |
| `bench_xsession.labels.jsonl` | `ec1431b9afd077cdc20026ca7886c5169f821ff4f0ec3224bba81a73d9af0b39` |

## Fragmentation types

F1 volatile values (timestamps, UUIDs, run ids) · F2 non-deterministic JSON serialization · F3 tool-set churn ·
F4 per-user interpolation · F5 conditional system sections · F6 whitespace/format drift · F7 retrieved context
inserted before stable instructions · F8 rewritten history.
At most one event per request/reference pair, so the cause of each first divergence is unambiguous.

## Benchmark headline

Measured on Kaggle 2 × Tesla T4 (free tier) with Qwen2.5-7B-Instruct-AWQ and vLLM 0.33.0, replaying these
traces at concurrency 1 (run `t4-20261111-0408`). Hit rate and prefill are determined by the cache algorithm,
not the GPU; TTFT/GPU-seconds (in `BENCH.md`) are **T4 reference values only**.

| trace | policy | token-weighted hit rate | recomputed prefill tokens |
|---|---|---:|---:|
| bench_main | raw | 74.71% | 241,640 |
| bench_main | L1 canonicalization | 81.79% | 172,268 (−28.7%) |
| bench_main | L2 relocation (reference) | 83.61% | 155,068 (−35.8%) |
| bench_xsession | raw | 86.63% | 50,120 |
| bench_xsession | L1 session-stable (deployed policy) | 86.69% | 49,273 |
| bench_xsession | L1 stateless name-sort (ceiling) | 95.31% | 17,369 |

L1/L2 are prompt transforms; identical generated text is not guaranteed (Phase 3 sampled L1-vs-raw match 10/18,
L2-vs-raw 4/6, against an 18/18 raw-vs-raw control). Full numbers, fidelity and the NVIDIA Dynamo trial are in
`BENCH.md`.

## Limitations

Synthetic text (word-salad sentences), synthetic tools, open-loop (assistant turns are generated, not model
outputs). Useful for measuring prefix-cache behaviour, not for training or evaluating language quality.