GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
singhmandavi
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
singhmandavi
Model Tree
Qwen/Qwen3-4B
Input Modalities
Output Modalities
Supported Functionality
python
from transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModelimport json base = "Qwen/Qwen3-4B"tok = AutoTokenizer.from_pretrained(base)model = AutoModelForCausalLM.from_pretrained(base, torch_dtype="auto", device_map="auto")model = PeftModel.from_pretrained(model, "singhmandavi/kg-slm-scierc-adapter") # prompt: system persona + SciERC ontology + 2 few-shot examples,# enable_thinking=False; greedy decoding; parse with salvage_json# (see kg_slm_phase1_scierc.ipynb, cells 1-4 for exact prompt/eval code)| metric | base Qwen3-4B | + adapter | + post-processing |
|---|---|---|---|
| schema validity | 0.937 | 1.000 | 1.000 |
| loose F1 | 0.070 | 0.438 | 0.441 |
| boundary (relaxed) F1 | 0.160 | 0.539 | 0.541 |
| Rel+ (published strict) | 0.046 | 0.411 | 0.414 |
| entity typed F1 | 0.316 | 0.666 | 0.669 |
| grounded rate | 0.977 | 0.997 | 1.000 |
Literature comparison: Rel+ 41.1 vs SpERT 38.6, SpERT.PL 39.4, PURE 36.8, PL-Marker 41.6; boundary 53.9 vs PL-Marker 53.2 (highest verified).
Single domain (CS abstracts), single seed, greedy decoding; entity F1 trails span-pipeline SOTA (~70 vs 66.6).
See REPORT.md (project report) — key refs: Luan et al. 2018 (SciERC),
Ye et al. 2022 (PL-Marker), Geng et al. 2023 (GCD), Dong et al. 2025
(XGrammar), Hu et al. 2021 (LoRA).
from transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModelimport json base = "Qwen/Qwen3-4B"tok = AutoTokenizer.from_pretrained(base)model = AutoModelForCausalLM.from_pretrained(base, torch_dtype="auto", device_map="auto")model = PeftModel.from_pretrained(model, "singhmandavi/kg-slm-scierc-adapter") # prompt: system persona + SciERC ontology + 2 few-shot examples,# enable_thinking=False; greedy decoding; parse with salvage_json# (see kg_slm_phase1_scierc.ipynb, cells 1-4 for exact prompt/eval code)