GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
itxmekan908067xcv
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
itxmekan908067xcv
Model Tree
Qwen/Qwen2.5-7B-Instruct
Input Modalities
Output Modalities
Supported Functionality
| Base model | Qwen2.5-7B-Instruct |
| Method | QLoRA (4-bit NF4 base, LoRA adapters) |
| LoRA | r = 16, alpha = 32, dropout = 0.05 |
| Target modules | q, k, v, o, gate, up, down projections |
| Trainable parameters | 40.4M (0.53% of 7.66B) |
| Language | English |
| License | Apache 2.0 (same as the base model) |
| Author | Atif Ullah Khan |
python
import torchfrom transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfigfrom peft import PeftModel BASE = "Qwen/Qwen2.5-7B-Instruct"ADAPTER = "YOUR_HF_USERNAME/cpc-pkl-qwen2.5-7b-lora" tok = AutoTokenizer.from_pretrained(ADAPTER)bnb = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4", bnb_4bit_use_double_quant=True, bnb_4bit_compute_dtype=torch.float16)model = AutoModelForCausalLM.from_pretrained(BASE, quantization_config=bnb, device_map="auto")model = PeftModel.from_pretrained(model, ADAPTER)model.eval() messages = [ {"role": "system", "content": "You are an expert Pakistani legal assistant specialising in the Code of Civil Procedure (CPC) 1908."}, {"role": "user", "content": "What is the procedural consequence when a plaint is returned under Order VII Rule 10 CPC?"},]inputs = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)with torch.no_grad(): out = model.generate(inputs, max_new_tokens=1024, do_sample=False, repetition_penalty=1.05)print(tok.decode(out[0][inputs.shape[1]:], skip_special_tokens=True))Use the same system prompt as above; the model was trained with it.
On 246 held-out test questions (greedy decoding):
| System | ROUGE-L | BERTScore-F1 | Provisions found (not named in question) | Grounded citations | Unverifiable citations |
|---|---|---|---|---|---|
| Base Qwen2.5-7B-Instruct | 0.234 | 0.841 | 3.0% | n/a | n/a |
| This adapter | 0.266 | 0.862 | 8.8% | 1.4% | 39.1% |
| This adapter + retrieval (RAG) | 0.278 | 0.864 | 23.8% | 64.9% | 22.6% |
Fine-tuning improves answer quality significantly over the base model (paired Wilcoxon, p < 1e-29), but it also teaches the model to cite cases it cannot ground. Combining the adapter with retrieval over source material greatly improves citation grounding; see the paper and code for the retrieval setup.
Research on legal NLP, hallucination and retrieval grounding for low-resource jurisdictions. Not for legal advice, court filings or any decision affecting someone's legal rights.
bibtex
@misc{khan2026cpcpkl, title = {CPC-PKL: Fine-Tuning, Retrieval, and Citation Hallucination in Legal Question Answering on Pakistan's Code of Civil Procedure}, author = {Khan, Atif Ullah}, year = {2026}, note = {Preprint. LINK_TO_ARXIV_OR_ZENODO}}import torchfrom transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfigfrom peft import PeftModel BASE = "Qwen/Qwen2.5-7B-Instruct"ADAPTER = "YOUR_HF_USERNAME/cpc-pkl-qwen2.5-7b-lora" tok = AutoTokenizer.from_pretrained(ADAPTER)bnb = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4", bnb_4bit_use_double_quant=True, bnb_4bit_compute_dtype=torch.float16)model = AutoModelForCausalLM.from_pretrained(BASE, quantization_config=bnb, device_map="auto")model = PeftModel.from_pretrained(model, ADAPTER)model.eval() messages = [ {"role": "system", "content": "You are an expert Pakistani legal assistant specialising in the Code of Civil Procedure (CPC) 1908."}, {"role": "user", "content": "What is the procedural consequence when a plaint is returned under Order VII Rule 10 CPC?"},]inputs = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)with torch.no_grad(): out = model.generate(inputs, max_new_tokens=1024, do_sample=False, repetition_penalty=1.05)print(tok.decode(out[0][inputs.shape[1]:], skip_special_tokens=True))@misc{khan2026cpcpkl, title = {CPC-PKL: Fine-Tuning, Retrieval, and Citation Hallucination in Legal Question Answering on Pakistan's Code of Civil Procedure}, author = {Khan, Atif Ullah}, year = {2026}, note = {Preprint. LINK_TO_ARXIV_OR_ZENODO}}