GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
kshetrajna12
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
kshetrajna12
Model Tree
Qwen/Qwen3.5-4B
Input Modalities
Output Modalities
Supported Functionality
bash
pip install uv && git clone https://github.com/kshetrajna12/reflex && cd reflex && uv syncuv run reflex-serve --model Qwen/Qwen3.5-4B --adapter kshetrajna12/reflex-qwen3.5-4b-lora --host 0.0.0.0 --port 8000
The adapter and its calibration.json are pulled from this repo automatically. Any CUDA GPU
with 16 GB or more works; the first request compiles kernels for about 20 s.
bash
curl -s localhost:8000/v1/systemone -H 'content-type: application/json' -d '{"state": {"ticket": "My payouts have failed three times this week and nobody replied."},"questions": {"queue": {"type": "choice", "instructions": "Which team should handle this?","criteria": {"payments": "payouts, refunds", "account": "login, 2FA", "other": null}},"escalate": {"type": "noul", "instructions": "Should this be escalated to a manager?"},"urgency": {"type": "score", "instructions": "How urgent is this?","criteria": ["can wait a week", "handle today", "blocked right now"]}}}'
json
{"answers": {"queue": {"type": "choice", "choice": "payments","probabilities": {"payments": 0.97, "account": 0.02, "other": 0.01}, "confidence": 0.91},"escalate": {"type": "noul", "noul": 0.71},"urgency": {"type": "score", "score": 1.77, "probabilities": {"0": 0.05, "1": 0.13, "2": 0.82},"legend": {"0": "can wait a week", "1": "handle today", "2": "blocked right now"}, "confidence": 0.62}},"usage": {"input_tokens": 258, "state_tokens": 79, "question_tokens": 179, "state_cache_hit": false}}
In Python, without a server:
python
from reflex import Engine, SystemOneRequestengine = Engine.load("Qwen/Qwen3.5-4B", adapter_path="kshetrajna12/reflex-qwen3.5-4b-lora") # calibration picked up from the reporesp = engine.answer(SystemOneRequest(state={"email": "I was charged twice this month, please fix."},questions={"team": {"type": "choice", "instructions": "Which team?", "criteria": {"billing": None, "technical": None, "sales": None}},"angry": {"type": "noul", "instructions": "Is the customer angry?"},},))resp.answers["team"].probabilities, resp.answers["angry"].noul
The request and response shapes are the same as TypeSafe's hosted /v1/systemone, so
Jev client code works against a reflex server unchanged.
| type | asks | you get back |
|---|---|---|
noul | "is this true?" | noul: probability of yes |
choice | "which one of these?" (up to 26 options) | choice, probabilities per option, confidence |
score | "how much, on this ordered scale?" (2 to 10 levels) | (probability-weighted level), , , |
state can be a string or any JSON, and can carry images as
{"type": "image", "source": "<path, URL or data: URI>"} (the base model is natively
multimodal). All questions in a request are answered independently, in parallel, over one
encoding of the state, so asking twenty questions about one document costs about the
same as asking one.
reflex-calibrate attaches LoRA adapters and minimises a proper scoring rule
(cross-entropy against a hard or soft target) on the model's label-restricted next-token
logits, the supervised form of Jev's "RLCD". Data came from reflex-data: public
datasets (banking77, CLINC, MMLU-Pro, civil_comments with soft toxicity labels, HaluEval,
MS MARCO, HelpSteer2 with per-rater score distributions, a GitHub code-review set) plus
synthetic, correct-by-construction data for instruction adequacy, rule-based routing and
reading probabilities off stated base rates. No benchmark items were used; an 8-gram
overlap check against JevBench's public items is part of the pipeline.
Fitted calibration temperature: 1.544 (applied automatically when served through reflex).
markdown
### alln=2082 acc=0.7963 ECE=0.0661 Brier=0.2985 NLL=0.6334 mean_conf=0.8507 fidelity=0.7589conf-bin mean_conf accuracy count[0.13,0.20) 0.175 0.000 3[0.20,0.27) 0.240 0.250 28[0.27,0.33) 0.294 0.206 34[0.33,0.40) 0.366 0.383 47[0.40,0.47) 0.435 0.345 58[0.47,0.53) 0.500 0.585 106[0.53,0.60) 0.568 0.593 86[0.60,0.67) 0.631 0.609 69[0.67,0.73) 0.701 0.628 78[0.73,0.80) 0.774 0.608 79[0.80,0.87) 0.836 0.732 97[0.87,0.93) 0.908 0.717 145[0.93,1.00) 0.988 0.942 1252### adequacy_synthn=100 acc=0.9400 ECE=0.0570 Brier=0.0971 NLL=0.2574 mean_conf=0.9767conf-bin mean_conf accuracy count[0.47,0.53) 0.531 0.000 2[0.53,0.60) 0.593 1.000 1[0.67,0.73) 0.679 1.000 1[0.73,0.80) 0.798 1.000 1[0.87,0.93) 0.915 1.000 1[0.93,1.00) 0.996 0.957 94### banking77n=200 acc=0.9350 ECE=0.0551 Brier=0.1170 NLL=0.3259 mean_conf=0.9857conf-bin mean_conf accuracy count[0.60,0.67) 0.658 0.500 2[0.73,0.80) 0.781 1.000 2[0.80,0.87) 0.828 0.667 3[0.87,0.93) 0.892 0.500 6[0.93,1.00) 0.997 0.957 187### base_rates_synthn=100 acc=0.9800 ECE=0.3625 Brier=0.2556 NLL=0.5043 mean_conf=0.6175 fidelity=0.9350conf-bin mean_conf accuracy count[0.33,0.40) 0.362 1.000 1[0.40,0.47) 0.446 0.750 8[0.47,0.53) 0.498 1.000 22[0.53,0.60) 0.568 1.000 21[0.60,0.67) 0.632 1.000 14[0.67,0.73) 0.699 1.000 13[0.73,0.80) 0.769 1.000 11[0.80,0.87) 0.831 1.000 7[0.87,0.93) 0.908 1.000 3### civil_commentsn=250 acc=0.9560 ECE=0.0440 Brier=0.0766 NLL=0.1458 mean_conf=0.9343 fidelity=0.8131conf-bin mean_conf accuracy count[0.47,0.53) 0.531 1.000 1[0.53,0.60) 0.577 0.500 2[0.60,0.67) 0.632 0.667 3[0.67,0.73) 0.700 1.000 5[0.73,0.80) 0.784 0.667 9[0.80,0.87) 0.835 0.917 12[0.87,0.93) 0.914 0.857 28[0.93,1.00) 0.968 0.995 190### clincn=200 acc=0.9350 ECE=0.0451 Brier=0.0976 NLL=0.2604 mean_conf=0.9778conf-bin mean_conf accuracy count[0.47,0.53) 0.500 0.500 2[0.53,0.60) 0.583 0.000 2[0.60,0.67) 0.634 0.000 1[0.67,0.73) 0.702 0.333 3[0.80,0.87) 0.839 0.333 3[0.87,0.93) 0.885 1.000 2[0.93,1.00) 0.997 0.973 187### codereviewn=320 acc=0.7750 ECE=0.1927 Brier=0.4177 NLL=1.0124 mean_conf=0.9591conf-bin mean_conf accuracy count[0.47,0.53) 0.500 1.000 1[0.53,0.60) 0.562 1.000 2[0.60,0.67) 0.637 0.500 2[0.67,0.73) 0.688 0.667 3[0.73,0.80) 0.777 0.000 1[0.80,0.87) 0.839 0.750 4[0.87,0.93) 0.916 0.611 54[0.93,1.00) 0.982 0.814 253### haluevaln=200 acc=0.9850 ECE=0.0149 Brier=0.0284 NLL=0.0695 mean_conf=0.9975conf-bin mean_conf accuracy count[0.87,0.93) 0.917 1.000 3[0.93,1.00) 0.999 0.985 197### helpsteer2n=125 acc=0.4400 ECE=0.1607 Brier=0.6694 NLL=1.3336 mean_conf=0.5753conf-bin mean_conf accuracy count[0.20,0.27) 0.251 1.000 1[0.27,0.33) 0.287 0.333 3[0.33,0.40) 0.356 0.500 4[0.40,0.47) 0.437 0.111 9[0.47,0.53) 0.497 0.367 30[0.53,0.60) 0.564 0.455 33[0.60,0.67) 0.623 0.444 18[0.67,0.73) 0.688 0.545 11[0.73,0.80) 0.780 0.400 5[0.80,0.87) 0.836 0.700 10[0.87,0.93) 0.877 1.000 1### helpsteer2_softn=125 acc=0.3680 ECE=0.1058 Brier=0.7793 NLL=1.5859 mean_conf=0.4103 fidelity=0.5902conf-bin mean_conf accuracy count[0.20,0.27) 0.248 0.375 8[0.27,0.33) 0.291 0.235 17[0.33,0.40) 0.365 0.371 35[0.40,0.47) 0.433 0.300 30[0.47,0.53) 0.496 0.591 22[0.53,0.60) 0.559 0.222 9[0.60,0.67) 0.647 0.000 2[0.67,0.73) 0.681 1.000 2### mmlu_pron=200 acc=0.5700 ECE=0.1429 Brier=0.5537 NLL=1.2712 mean_conf=0.6951conf-bin mean_conf accuracy count[0.13,0.20) 0.175 0.000 3[0.20,0.27) 0.237 0.158 19[0.27,0.33) 0.299 0.143 14[0.33,0.40) 0.374 0.286 7[0.40,0.47) 0.433 0.364 11[0.47,0.53) 0.496 0.312 16[0.53,0.60) 0.556 1.000 4[0.60,0.67) 0.620 0.400 10[0.67,0.73) 0.710 0.556 9[0.73,0.80) 0.764 0.500 10[0.80,0.87) 0.844 0.667 12[0.87,0.93) 0.898 0.500 18[0.93,1.00) 0.981 0.940 67### msmarcon=200 acc=0.6550 ECE=0.1583 Brier=0.4644 NLL=0.6593 mean_conf=0.7742conf-bin mean_conf accuracy count[0.47,0.53) 0.522 0.800 10[0.53,0.60) 0.583 0.417 12[0.60,0.67) 0.639 0.706 17[0.67,0.73) 0.707 0.452 31[0.73,0.80) 0.775 0.525 40[0.80,0.87) 0.836 0.696 46[0.87,0.93) 0.897 0.862 29[0.93,1.00) 0.960 0.933 15### rule_routing_synthn=62 acc=1.0000 ECE=0.0007 Brier=0.0000 NLL=0.0007 mean_conf=0.9993conf-bin mean_conf accuracy count[0.93,1.00) 0.999 1.000 62
scoreprobabilitieslegendconfidence