GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Gydeai
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Gydeai
Model Tree
Qwen/Qwen3-8B
Input Modalities
Output Modalities
Supported Functionality
The 2,000 newest single-component complaints received 6 to 28 October 2026, all newer than every training complaint (training: January 2025 to 25 September 2026). Same test set for every model.
| Model | Accuracy | Macro-F1 | Invalid answers |
|---|---|---|---|
| This adapter (Qwen3-8B + LoRA, rank 16) | 87.6% | 0.846 | 0 |
| Claude Sonnet 5.5, component list in the prompt | 86.3% | 0.828 | 0 |
| Claude Haiku 5.5, component list in the prompt | 85.6% | 0.810 | 1 |
| TF-IDF keyword classifier, same training data | 82.2% | ||
| Qwen3-8B untuned, component list in the prompt | 77.5% | 0.729 | 3 |
The adapter and Claude Sonnet are within about one percentage point of each other on this set; read it as a match, not a lead. Fine-tuning added about 10 points over the untuned base model.
| Model | Real-time | Nightly batch of 1,000 | Batch of 10,000 |
|---|---|---|---|
| This adapter on one NVIDIA A30 | ~$0.004 | ~$0.022 | ~$0.006 |
| This adapter on one NVIDIA L4 | ~$0.007 | ~$0.024 | ~$0.009 |
| Claude Haiku 5.5 | ~$0.11 | ~$0.056 | ~$0.056 |
| Claude Sonnet 5.5 | ~$1.68 | ~$0.84 | ~$0.84 |
Base model Qwen/Qwen3-8B (revision b968826d9c46dd6066d109eabc6255188de91218). Greedy decoding (temperature 0),
thinking disabled. The adapter was trained on exactly this prompt:
You route US vehicle owners' defect complaints to the component team. Reply with JSON only: {"component": ...}.Complaint: {text}\nAnswer:With vLLM:
bash
vllm serve Qwen/Qwen3-8B --revision b968826d9c46dd6066d109eabc6255188de91218 \ --enable-lora --max-lora-rank 16 --lora-modules us-nhtsa=Gydeai/vehicle-complaint-triage-us-nhtsapython
from openai import OpenAIclient = OpenAI(base_url="http://localhost:8000/v1", api_key="local")r = client.chat.completions.create( model="us-nhtsa", temperature=0, max_tokens=40, extra_body={"chat_template_kwargs": {"enable_thinking": False}}, messages=[{"role": "system", "content": "You route US vehicle owners' defect complaints to the component team. " 'Reply with JSON only: {"component": ...}.'}, {"role": "user", "content": "Complaint: The brake pedal went to the floor at a stop light and the car " "rolled into the intersection.\nAnswer:"}])print(r.choices[0].message.content) # e.g. {"component": "SERVICE BRAKES"}labels.json lists the 17 component groups.
Gyde trains versions on your own categories (warranty codes, service-ticket queues) and runs them in your cloud: gyde.ai.
LoRA rank 16, alpha 32, dropout 0.05 on all attention and MLP projections; 1 epoch over 8,000 complaints, learning
rate 2e-4; 20 minutes on one A100 80 GB (run.json). The loss covered only the answer (the component label), not
the complaint text.
NHTSA Office of Defects Investigation complaints, file COMPLAINTS_RECEIVED_2025-2026 from
static.nhtsa.gov/odi/ffdd/cmpl (sha256 in labels.json), a US
government work. Kept: complaints filed under exactly one component, narratives of 200 to 4,000 characters, with
"Unknown or other" dropped, renamed groups merged (for example "Service brakes, hydraulic" into "Service brakes"),
groups with fewer than 400 complaints dropped, and recall-campaign texts that quote the component removed.
Split by date: 8,000 training (from older complaints), 1,000 validation, 2,000 test (the newest).
The training data is not distributed with the adapter. NHTSA blanks most personal details; a few narratives still contain a phone number or VIN. The adapter was trained only to output a component label.
Adapter: Apache 2.0. Base model: Qwen3-8B under its own licence (Apache 2.0). Data: NHTSA, a US government work.
Built by Gyde (DAI Labs Private Limited, Pune, India).
vllm serve Qwen/Qwen3-8B --revision b968826d9c46dd6066d109eabc6255188de91218 \ --enable-lora --max-lora-rank 16 --lora-modules us-nhtsa=Gydeai/vehicle-complaint-triage-us-nhtsafrom openai import OpenAIclient = OpenAI(base_url="http://localhost:8000/v1", api_key="local")r = client.chat.completions.create( model="us-nhtsa", temperature=0, max_tokens=40, extra_body={"chat_template_kwargs": {"enable_thinking": False}}, messages=[{"role": "system", "content": "You route US vehicle owners' defect complaints to the component team. " 'Reply with JSON only: {"component": ...}.'}, {"role": "user", "content": "Complaint: The brake pedal went to the floor at a stop light and the car " "rolled into the intersection.\nAnswer:"}])print(r.choices[0].message.content) # e.g. {"component": "SERVICE BRAKES"}