Datasets:
File size: 5,094 Bytes
0907996 4432332 4ae284e 4432332 0907996 4432332 0907996 4432332 0907996 4432332 0907996 4432332 0907996 4ae284e 0907996 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 | ---
license: apache-2.0
pretty_name: agentx-mini
task_categories:
- text-generation
tags:
- llm
- inference
- agentic
- benchmarking
- kv-cache
- multi-turn
- subagents
size_categories:
- n<1K
configs:
- config_name: default
data_files:
- split: train
path: traces.jsonl
---
# agentx-mini
**agentx-mini** is a smaller workload for benchmarking local LLM servers, derived from
[semianalysisai/cc-traces-weka-062126-256k](https://huggingface.co/datasets/semianalysisai/cc-traces-weka-062126-256k).
It contains **200 sessions and 9,439 model calls**, including **1,363 subagent calls**.
Changes from the original dataset:
- Divide input and output lengths by **8**, rounding each output up. Hash IDs stay the same; logical blocks shrink from 64 to 8 tokens.
- Divide active request time by **8** and shorten idle gaps. Replay uses a **1-second idle cap**.
- Keep complete sessions with **less than 10,000 requested output tokens**, counting all subagents and auxiliary calls.
- Preserve request grouping, spawn/join dependencies, shared prompt content, and reference overlaps. Actual replay overlap depends on server speed.
## Stats
```text
traces: 200
main_turns: 8,076
subagent_groups: 76
subagent_inner_requests: 1,363
total_model_requests: 9,439
total_input_tokens: 126,413,728
total_output_tokens: 962,854
```
Main turns are top-level calls. Input totals include repeated prefixes; actual
prefill depends on the server's prefix caching. Rendered prompts add
tokenizer/chat-template overhead.
## Files and reproduction
`traces.jsonl` contains the scaled Weka records; `stats.txt` contains their totals.
AgentX Harness handles prompt synthesis and replay using the target model's tokenizer.
To rebuild agentx-mini using [distill.py](distill.py):
```sh
pip install huggingface-hub orjson
python distill.py --output rebuilt
```
## Run agentx-mini with AgentX Harness
Requires Python 3.11–3.13. Install in a dedicated virtual environment, following
the [AgentX Harness README](https://github.com/SemiAnalysisAI/agentx-harness#install-and-run-agentx):
```bash
python3 -m venv .venv
source .venv/bin/activate
pip install agentx-harness==1.0.7
```
**Set the block size for agentx-mini.** AgentX 1.0.7 needs an adjustment for this
dataset's 8-token blocks. In the activated virtual environment, locate the
installed plugin configuration:
```bash
python -c 'from pathlib import Path; import aiperf.plugin; print(Path(aiperf.plugin.__file__).with_name("plugins.yaml"))'
```
Open that file. Under `public_dataset_loader` → `weka_hf` → `metadata`, change
`default_block_size: 64` to `default_block_size: 8`. This setting applies to all
`weka_hf` datasets in this environment; reinstalling the harness can reset it.
1. **Download agentx-mini** into a local directory.
2. **Start your local LLM server** and load your model, for example using
`llama-server`. It must expose an OpenAI-compatible API. Enable parallel
request slots to exercise subagents. Allow up to 32,000 tokens per request
for input plus output, with additional room for chat-template overhead.
Keep the server running during the benchmark.
3. **Configure AgentX.** Save the configuration below as `benchmark.yaml` beside
`traces.jsonl`. Replace the served model name, tokenizer, and server URL with
your server's settings. Set `hf_weka_dataset` to the **absolute path** of the
downloaded dataset directory (run `pwd` there to get it).
```yaml
random_seed: 42
benchmark:
models: [YOUR_SERVED_MODEL_NAME]
tokenizer:
name: YOUR_TOKENIZER
endpoint:
urls: [http://127.0.0.1:8000]
streaming: true
extra: {ignore_eos: true}
use_legacy_max_tokens: true
datasets:
- name: default
type: public
dataset: weka_hf
hf_weka_dataset: /absolute/path/to/agentx-mini
sampling: sequential
trace_idle_gap_cap_seconds: 1
phases:
- name: profiling
type: concurrency
timing_mode: agentic_replay
concurrency: 1
duration: 1200
trajectory_start_min_ratio: 0
trajectory_start_max_ratio: 0
```
Run from that directory:
```bash
agentx profile \
--config benchmark.yaml \
--output-artifact-dir ./results/agentx-mini
```
**Automatic download:** instead of downloading first, set `hf_weka_dataset` to
`SocialLocalMobile/agentx-mini` and run the same command.
## License
[Apache 2.0](LICENSE), matching the source dataset published by SemiAnalysis
from its Claude Code proxy traces. The dataset, script, and documentation in
this repository are distributed under these terms.
**Modified files:** `traces.jsonl` is a scaled and filtered derivative of source
revision [`8fecd2f`](https://huggingface.co/datasets/semianalysisai/cc-traces-weka-062126-256k/tree/8fecd2fc56694469f758f0afbbb6335ad3043740);
`stats.txt` is regenerated, and this README and `.gitattributes` are adapted for
agentx-mini. The data changes are listed above. References to SemiAnalysis
identify the source and do not imply endorsement.
|