GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Calm153
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Calm153
Model Tree
Qwen/Qwen3-1.7B
Input Modalities
Output Modalities
Supported Functionality
Fresh base-versus-tuned rerun on the same frozen 30 conversations / 180 turns:
| Model | Exact Spec adherence | Robustness | Full-conversation pass |
|---|---|---|---|
| Base | 1.7% (3/180) | 2.9% (3/102) | 0% |
| N=128 | 70.6% (127/180) | 64.7% (66/102) | 53.3% (16/30) |
Both models used the same structured prompt, greedy/no-thinking decoding, max 128 new tokens, and NF4 inference. The base emitted invalid JSON on 140/180 turns. All 360 informed judge verdicts agreed with the deterministic simulator; the judge sees the oracle and explains its result, so it is not an independent correctness oracle. Full scores, reasoning, raw responses, generations, source, and hashes are in evidence/.
This evaluation set previously informed checkpoint selection. Robustness is a 102-turn rejection-recovery subset. No staff scenarios exist (the user clarified a PRD oversight). The 100% reliability target is unmet; do not claim a minimum viable dataset size or general frontier-model superiority. This fictional benchmark is not suitable for industrial control or real safety decisions.
Use a CUDA GPU and the pinned versions in evidence/code/requirements.txt.
python
import torchfrom transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfigfrom peft import PeftModel base = "Qwen/Qwen3-1.7B"base_revision = "70d244cc86ccca08cf5af4e1e306ecf908b1ad5e"# Replace with the exact published adapter commit from release.json/submission.adapter_revision = "d76a5c3c7d947653f6cbaabf4c85030cc610fa40"tokenizer = AutoTokenizer.from_pretrained(base, revision=base_revision)model = AutoModelForCausalLM.from_pretrained( base, revision=base_revision, device_map="auto", quantization_config=BitsAndBytesConfig( load_in_4bit=True, bnb_4bit_quant_type="nf4", bnb_4bit_use_double_quant=True, bnb_4bit_compute_dtype=torch.float16))model = PeftModel.from_pretrained( model, "Calm153/vitrion-qwen3-1.7b-qlora-n128", revision=adapter_revision).eval()The model is released as a PEFT adapter; its public base weights are downloaded separately. Use apply_chat_template(..., enable_thinking=False), a system prompt from the frozen dataset.strategies()['structured'], and generate(max_new_tokens=128, do_sample=False). Only the first user turn should contain the initial state. Append the model's actual responses to its history, including mistakes.
import torchfrom transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfigfrom peft import PeftModel base = "Qwen/Qwen3-1.7B"base_revision = "70d244cc86ccca08cf5af4e1e306ecf908b1ad5e"# Replace with the exact published adapter commit from release.json/submission.adapter_revision = "d76a5c3c7d947653f6cbaabf4c85030cc610fa40"tokenizer = AutoTokenizer.from_pretrained(base, revision=base_revision)model = AutoModelForCausalLM.from_pretrained( base, revision=base_revision, device_map="auto", quantization_config=BitsAndBytesConfig( load_in_4bit=True, bnb_4bit_quant_type="nf4", bnb_4bit_use_double_quant=True, bnb_4bit_compute_dtype=torch.float16))model = PeftModel.from_pretrained( model, "Calm153/vitrion-qwen3-1.7b-qlora-n128", revision=adapter_revision).eval()