GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
pyaroslav
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
pyaroslav
Model Tree
Qwen/Qwen3-8B
Input Modalities
Output Modalities
Supported Functionality
| Qwen3-8B | AWR | Statspack | Top cause right | False causes |
|---|---|---|---|---|
| Untuned | 5/9 | 6/9 | 11/18 | 38 |
| This adapter | 8/9 | 8/9 | 18/18 | 0 |
A run passes when the injected cause is in the top 2, or both causes are in the top 3 when two were injected. cpu_cap isn't counted as a false cause. Both misses are a full-scan + commit-storm pair where the model names only the commit storm.
These are lab faults, one or two at a time, on short windows. Real systems are messier; treat the output as a first pass for a DBA, not a verdict.
adapter_model.safetensors, adapter_config.json: PEFT format for Qwen/Qwen3-8B.second-opinion-qwen3-lora-f16.gguf: the same adapter for llama.cpp. In my tests it works when merged with llama-export-lora into Ollama's qwen3:8b Q4_K_M weights and quantized to Q8_0 (lab/tune/export.py). There it agreed with the Tinker-hosted model's top cause on 17 of 18 held-out runs. Requantizing to Q4 lost the tuning.python
from peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizer base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-8B", torch_dtype="auto", device_map="auto")model = PeftModel.from_pretrained(base, "pyaroslav/second-opinion-qwen3-8b-lora")tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-8B")from peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizer base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-8B", torch_dtype="auto", device_map="auto")model = PeftModel.from_pretrained(base, "pyaroslav/second-opinion-qwen3-8b-lora")tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-8B")