zip-glm47-sy
#16
by snehalyelmati - opened
results/zip-glm47-sy/README.md
ADDED
|
@@ -0,0 +1,40 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
### zip-glm47-sy
|
| 2 |
+
|
| 3 |
+
Zip is a lightweight browser-use agent runtime for LLM-driven web automation. It runs as a BrowserGym-compatible AgentLab agent for benchmark evaluation.
|
| 4 |
+
|
| 5 |
+
#### Model Name
|
| 6 |
+
|
| 7 |
+
- `z-ai/glm-4.7:nitro` via OpenRouter
|
| 8 |
+
|
| 9 |
+
#### Model Architecture
|
| 10 |
+
|
| 11 |
+
Zip uses a general-purpose browser-agent architecture with structured browser context extraction, stable element IDs, semantic DOM-first tools, and an optional planner/worker split. This MiniWoB run used unified mode, where a single tool-equipped agent acts after context extraction, filtering, and oracle checks.
|
| 12 |
+
|
| 13 |
+
#### Input/Output Format
|
| 14 |
+
|
| 15 |
+
The agent receives BrowserGym task instructions plus a pruned structured page snapshot containing element IDs, labels, text context, attributes, handler hints, and recent state. It outputs semantic browser tool calls such as click, type, select, read, scroll, drag, and completion decisions.
|
| 16 |
+
|
| 17 |
+
#### Training Details
|
| 18 |
+
|
| 19 |
+
No agent-specific training or fine-tuning was used. The run used the base hosted model with Zip's runtime, prompts, tools, and benchmark configuration.
|
| 20 |
+
|
| 21 |
+
#### Benchmark Configuration
|
| 22 |
+
|
| 23 |
+
```bash
|
| 24 |
+
uv run --extra agentlab python benchmarks/agentlab/run_browsergym_benchmark.py \
|
| 25 |
+
--benchmark miniwob \
|
| 26 |
+
--preset full \
|
| 27 |
+
--n-repeats 1 \
|
| 28 |
+
--max-steps 20 \
|
| 29 |
+
--env-max-steps 10 \
|
| 30 |
+
--max-elements 80 \
|
| 31 |
+
--n-jobs 1
|
| 32 |
+
```
|
| 33 |
+
|
| 34 |
+
#### Additional Notes
|
| 35 |
+
|
| 36 |
+
- Benchmark: MiniWoB full suite
|
| 37 |
+
- Episodes: 125
|
| 38 |
+
- Study ID: `2026-05-26_15-41-13`
|
| 39 |
+
- Provider: OpenRouter
|
| 40 |
+
- Model: `z-ai/glm-4.7:nitro`
|
results/zip-glm47-sy/miniwob.json
ADDED
|
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"agent_name": "zip-glm47-sy",
|
| 4 |
+
"study_id": "2026-05-26_15-41-13",
|
| 5 |
+
"date_time": "2026-05-26 15:41:13",
|
| 6 |
+
"benchmark": "MiniWoB",
|
| 7 |
+
"score": 86.4,
|
| 8 |
+
"std_err": 3.066,
|
| 9 |
+
"benchmark_specific": "No",
|
| 10 |
+
"benchmark_tuned": "No",
|
| 11 |
+
"followed_evaluation_protocol": "Yes",
|
| 12 |
+
"reproducible": "Yes",
|
| 13 |
+
"comments": "Zip unified-mode browser agent using OpenRouter z-ai/glm-4.7:nitro; MiniWoB full suite, 125 episodes, one repeat, max_steps=20, env_max_steps=10, max_elements=80.",
|
| 14 |
+
"original_or_reproduced": "Original"
|
| 15 |
+
}
|
| 16 |
+
]
|