GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
teacher57
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
teacher57
Model Tree
unsloth/Qwen3-14B-unsloth-bnb-4bit
Input Modalities
Output Modalities
Supported Functionality
| retail test, 115 tasks | this adapter | starting adapter, same-pod re-run | starting adapter, earlier test (2 trials, other pod) |
|---|---|---|---|
| pass^1 | 45.7% ± 3.4 | 41.7% ± 3.4 | 37.8% ± 3.7 |
| pass^2 | 29.9% | 27.1% | 21.7% |
| pass^3 | 22.0% | 20.2% | n/a |
| pass^4 | 16.5% | 16.5% | n/a |
By task type (pass^1): combo tasks (64) 34.8% vs 33.6%; other tasks (51) 59.3% vs 52.0%. That split was made after seeing the results, so it is a hypothesis, not a finding. No airline result exists (that test was stopped after 10 rollouts).
<think> reasoning, the tool call or reply and <|im_end|>).
python
from peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizerbase = "unsloth/Qwen3-14B-unsloth-bnb-4bit"model = AutoModelForCausalLM.from_pretrained(base, device_map="auto")model = PeftModel.from_pretrained(model, "teacher57/qwen3-14b-tau-distilled-from-32b")tok = AutoTokenizer.from_pretrained("teacher57/qwen3-14b-tau-distilled-from-32b")For τ-bench evaluation we served it with vLLM 0.11.0 (--enable-lora --max-loras 3 --max-lora-rank 16, tool parser hermes,
reasoning parser qwen3).
Retail domain only (the airline test was not completed); trained and tested with GPT-4o as the simulated customer; the training conversations come from one teacher and one benchmark, so nothing here says the adapter generalises beyond τ-bench retail.
from peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizerbase = "unsloth/Qwen3-14B-unsloth-bnb-4bit"model = AutoModelForCausalLM.from_pretrained(base, device_map="auto")model = PeftModel.from_pretrained(model, "teacher57/qwen3-14b-tau-distilled-from-32b")tok = AutoTokenizer.from_pretrained("teacher57/qwen3-14b-tau-distilled-from-32b")