Datasets:
Clarify AgentX setup and simplify README
Browse files
README.md
CHANGED
|
@@ -25,14 +25,13 @@ configs:
|
|
| 25 |
**agentx-mini** is a smaller workload for benchmarking local LLM servers, derived from
|
| 26 |
[semianalysisai/cc-traces-weka-062126-256k](https://huggingface.co/datasets/semianalysisai/cc-traces-weka-062126-256k).
|
| 27 |
It contains **200 sessions and 9,439 model calls**, including **1,363 subagent calls**.
|
| 28 |
-
The original dataset's filters still apply; the changes below are additional.
|
| 29 |
|
| 30 |
Changes from the original dataset:
|
| 31 |
|
| 32 |
- Divide input and output lengths by **8**, rounding each output up. Hash IDs stay the same; logical blocks shrink from 64 to 8 tokens.
|
| 33 |
-
- Divide active request time by **8** and
|
| 34 |
- Keep complete sessions with **less than 10,000 requested output tokens**, counting all subagents and auxiliary calls.
|
| 35 |
-
- Preserve request grouping, spawn/join dependencies, shared prompt content, and reference overlaps.
|
| 36 |
|
| 37 |
## Stats
|
| 38 |
|
|
@@ -46,29 +45,22 @@ total_input_tokens: 126,413,728
|
|
| 46 |
total_output_tokens: 962,854
|
| 47 |
```
|
| 48 |
|
| 49 |
-
Main turns are top-level calls
|
| 50 |
-
|
| 51 |
-
|
| 52 |
-
|
| 53 |
-
Input totals count repeated prefixes. The ideal estimate is **5,499,568 uncached
|
| 54 |
-
input tokens** (**95.65% prefix reuse**), assuming a cold, unlimited cache shared
|
| 55 |
-
by each session's agents on one model. Actual prefill depends on the server.
|
| 56 |
|
| 57 |
## Files and reproduction
|
| 58 |
|
| 59 |
`traces.jsonl` contains the scaled Weka records; `stats.txt` contains their totals.
|
| 60 |
AgentX Harness handles prompt synthesis and replay using the target model's tokenizer.
|
| 61 |
|
| 62 |
-
To rebuild agentx-mini:
|
| 63 |
|
| 64 |
```sh
|
| 65 |
pip install huggingface-hub orjson
|
| 66 |
python distill.py --output rebuilt
|
| 67 |
```
|
| 68 |
|
| 69 |
-
[distill.py](distill.py) pins the source revision and applies deterministic token
|
| 70 |
-
and timeline transforms. It needs no model, tokenizer, or replay plugin.
|
| 71 |
-
|
| 72 |
## Run agentx-mini with AgentX Harness
|
| 73 |
|
| 74 |
Requires Python 3.11–3.13. Install in a dedicated virtual environment, following
|
|
@@ -80,6 +72,18 @@ source .venv/bin/activate
|
|
| 80 |
pip install agentx-harness==1.0.7
|
| 81 |
```
|
| 82 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 83 |
1. **Download agentx-mini** into a local directory.
|
| 84 |
2. **Start your local LLM server** and load your model, for example using
|
| 85 |
`llama-server`. It must expose an OpenAI-compatible API. Enable parallel
|
|
@@ -92,6 +96,7 @@ pip install agentx-harness==1.0.7
|
|
| 92 |
downloaded dataset directory (run `pwd` there to get it).
|
| 93 |
|
| 94 |
```yaml
|
|
|
|
| 95 |
benchmark:
|
| 96 |
models: [YOUR_SERVED_MODEL_NAME]
|
| 97 |
tokenizer:
|
|
@@ -123,19 +128,12 @@ Run from that directory:
|
|
| 123 |
```bash
|
| 124 |
agentx profile \
|
| 125 |
--config benchmark.yaml \
|
| 126 |
-
--random-seed 42 \
|
| 127 |
--output-artifact-dir ./results/agentx-mini
|
| 128 |
```
|
| 129 |
|
| 130 |
**Automatic download:** instead of downloading first, set `hf_weka_dataset` to
|
| 131 |
`SocialLocalMobile/agentx-mini` and run the same command.
|
| 132 |
|
| 133 |
-
Keep `timing_mode: agentic_replay` and the runtime idle cap to skip retained long
|
| 134 |
-
reference gaps. The harness can occasionally exceed the idle cap around joins.
|
| 135 |
-
|
| 136 |
-
Call coverage, prefix reuse, and dependencies were validated across all 200
|
| 137 |
-
sessions. Also tested with `agentx-harness==1.0.7` and Qwen3-0.6B on `llama-server`.
|
| 138 |
-
|
| 139 |
## License
|
| 140 |
|
| 141 |
[Apache 2.0](LICENSE), matching the source dataset published by SemiAnalysis
|
|
|
|
| 25 |
**agentx-mini** is a smaller workload for benchmarking local LLM servers, derived from
|
| 26 |
[semianalysisai/cc-traces-weka-062126-256k](https://huggingface.co/datasets/semianalysisai/cc-traces-weka-062126-256k).
|
| 27 |
It contains **200 sessions and 9,439 model calls**, including **1,363 subagent calls**.
|
|
|
|
| 28 |
|
| 29 |
Changes from the original dataset:
|
| 30 |
|
| 31 |
- Divide input and output lengths by **8**, rounding each output up. Hash IDs stay the same; logical blocks shrink from 64 to 8 tokens.
|
| 32 |
+
- Divide active request time by **8** and shorten idle gaps. Replay uses a **1-second idle cap**.
|
| 33 |
- Keep complete sessions with **less than 10,000 requested output tokens**, counting all subagents and auxiliary calls.
|
| 34 |
+
- Preserve request grouping, spawn/join dependencies, shared prompt content, and reference overlaps. Actual replay overlap depends on server speed.
|
| 35 |
|
| 36 |
## Stats
|
| 37 |
|
|
|
|
| 45 |
total_output_tokens: 962,854
|
| 46 |
```
|
| 47 |
|
| 48 |
+
Main turns are top-level calls. Input totals include repeated prefixes; actual
|
| 49 |
+
prefill depends on the server's prefix caching. Rendered prompts add
|
| 50 |
+
tokenizer/chat-template overhead.
|
|
|
|
|
|
|
|
|
|
|
|
|
| 51 |
|
| 52 |
## Files and reproduction
|
| 53 |
|
| 54 |
`traces.jsonl` contains the scaled Weka records; `stats.txt` contains their totals.
|
| 55 |
AgentX Harness handles prompt synthesis and replay using the target model's tokenizer.
|
| 56 |
|
| 57 |
+
To rebuild agentx-mini using [distill.py](distill.py):
|
| 58 |
|
| 59 |
```sh
|
| 60 |
pip install huggingface-hub orjson
|
| 61 |
python distill.py --output rebuilt
|
| 62 |
```
|
| 63 |
|
|
|
|
|
|
|
|
|
|
| 64 |
## Run agentx-mini with AgentX Harness
|
| 65 |
|
| 66 |
Requires Python 3.11–3.13. Install in a dedicated virtual environment, following
|
|
|
|
| 72 |
pip install agentx-harness==1.0.7
|
| 73 |
```
|
| 74 |
|
| 75 |
+
**Set the block size for agentx-mini.** AgentX 1.0.7 needs an adjustment for this
|
| 76 |
+
dataset's 8-token blocks. In the activated virtual environment, locate the
|
| 77 |
+
installed plugin configuration:
|
| 78 |
+
|
| 79 |
+
```bash
|
| 80 |
+
python -c 'from pathlib import Path; import aiperf.plugin; print(Path(aiperf.plugin.__file__).with_name("plugins.yaml"))'
|
| 81 |
+
```
|
| 82 |
+
|
| 83 |
+
Open that file. Under `public_dataset_loader` → `weka_hf` → `metadata`, change
|
| 84 |
+
`default_block_size: 64` to `default_block_size: 8`. This setting applies to all
|
| 85 |
+
`weka_hf` datasets in this environment; reinstalling the harness can reset it.
|
| 86 |
+
|
| 87 |
1. **Download agentx-mini** into a local directory.
|
| 88 |
2. **Start your local LLM server** and load your model, for example using
|
| 89 |
`llama-server`. It must expose an OpenAI-compatible API. Enable parallel
|
|
|
|
| 96 |
downloaded dataset directory (run `pwd` there to get it).
|
| 97 |
|
| 98 |
```yaml
|
| 99 |
+
random_seed: 42
|
| 100 |
benchmark:
|
| 101 |
models: [YOUR_SERVED_MODEL_NAME]
|
| 102 |
tokenizer:
|
|
|
|
| 128 |
```bash
|
| 129 |
agentx profile \
|
| 130 |
--config benchmark.yaml \
|
|
|
|
| 131 |
--output-artifact-dir ./results/agentx-mini
|
| 132 |
```
|
| 133 |
|
| 134 |
**Automatic download:** instead of downloading first, set `hf_weka_dataset` to
|
| 135 |
`SocialLocalMobile/agentx-mini` and run the same command.
|
| 136 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 137 |
## License
|
| 138 |
|
| 139 |
[Apache 2.0](LICENSE), matching the source dataset published by SemiAnalysis
|