GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
kushagrasoni6446
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
kushagrasoni6446
Model Tree
Qwen/Qwen3-1.7B
Input Modalities
Output Modalities
Supported Functionality
python
from peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizer base = "Qwen/Qwen3-1.7B"tok = AutoTokenizer.from_pretrained(base)model = PeftModel.from_pretrained(AutoModelForCausalLM.from_pretrained(base, dtype="auto"), "kushagrasoni6446/qwen3-1.7b-invoice-extraction-lora")SYSTEM = open("prompts_system.txt").read() # the exact system prompt is in this repo; use it unchangedmsgs = [{"role": "system", "content": SYSTEM}, {"role": "user", "content": "Document:\n" + invoice_text}]prompt = tok.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True, enable_thinking=False)out = model.generate(**tok(prompt, return_tensors="pt"), max_new_tokens=700, do_sample=False)The adapter was trained on exactly this prompt format (the chat template, enable_thinking=False, greedy
decoding). Other prompts are untested.
Test: 300 invoices, half in 2 layouts held out of training. Shift: 100 invoices in 2 other layouts, read by Tesseract from scans and degraded scans. "Fully correct" means every header field and every line item is right.
| Test: fully correct | Header fields | Shift: fully correct | Invented values / invoice (shift) | |
|---|---|---|---|---|
| this adapter (seed 9, chosen on dev) | 80.7% | 0.984 | 27% | 0.39 |
| same recipe, 3 seeds | 79.0% (77.7–80.7) | 0.978 | 28.3% (27–30) | 0.29 |
| Qwen3.6-35B-A3B, 3 examples | 76.3% | 0.988 | 51% | 0.00 |
| Qwen3-1.7B, 3 examples, no adapter | 57.7% | 0.958 | 44% | 0.07 |
| Qwen3-1.7B, zero-shot | 29.3% | 0.897 | 36% | 0.09 |
Training: rank 64, alpha 128, all linear layers, lr 2e-4 (cosine), 1 epoch, 4,000 documents, completion-only loss. It took 54 minutes on one Kaggle T4 (fp16), with a 5.3 GB training peak and about 6 GB serving at batch 16.
Fully synthetic: company names and addresses from Faker, generated by code in the project. No real person, company or document. See the data card there.
Apache-2.0, the same as the base model (Qwen/Qwen3-1.7B).
from peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizer base = "Qwen/Qwen3-1.7B"tok = AutoTokenizer.from_pretrained(base)model = PeftModel.from_pretrained(AutoModelForCausalLM.from_pretrained(base, dtype="auto"), "kushagrasoni6446/qwen3-1.7b-invoice-extraction-lora")SYSTEM = open("prompts_system.txt").read() # the exact system prompt is in this repo; use it unchangedmsgs = [{"role": "system", "content": SYSTEM}, {"role": "user", "content": "Document:\n" + invoice_text}]prompt = tok.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True, enable_thinking=False)out = model.generate(**tok(prompt, return_tensors="pt"), max_new_tokens=700, do_sample=False)