GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
bharathcherala193
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
bharathcherala193
Model Tree
Qwen/Qwen2.5-1.5B-Instruct
Input Modalities
Output Modalities
Supported Functionality
The prompt must be built with the 12-tool schema and the training system prompt:
You are Aria, a helpful e-commerce assistant. You help customers search for products, track orders, manage returns, answer questions, and resolve issues. Always be polite, concise, and accurate.
utils.py in this repo contains TOOLS_SCHEMA, build_chat_prompt_text
(applies the chat template with the tools) and parse_model_output.
Load the base model in 4-bit and attach this adapter with PeftModel.from_pretrained.
Held-out subset of 300 test examples (120 tool-call, 90 missing-info, 90 no-tool), greedy decoding.
| Metric | Base model | This adapter |
|---|---|---|
| valid_format_rate | 0.05 | 1.00 |
| exact_match_rate | 0.003 | 0.643 |
| tool_name_accuracy | 0.083 | 0.625 |
| argument_accuracy | 0.008 | 0.308 |
| no_tool_accuracy | 0.00 | 0.867 |
| missing_info_accuracy | 0.00 | 0.867 |
NO_TOOL_CALL on simple search or add-to-cart requests.QLoRA (4-bit NF4), r=16, alpha=32, 300 steps, learning rate 2e-4, on 5,541 filtered examples from Full-Ecom-Chatbot-Dataset. Mixed training targets: 1,022 tool calls, 2,000 missing-info, 3,000 no-tool (the last two capped for balance).