GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
ManideepVarma003
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
ManideepVarma003
Model Tree
unsloth/Qwen3-4B-Instruct-2507-unsloth-bnb-4bit
Input Modalities
Output Modalities
Supported Functionality
Requires a CUDA GPU (the base model is a 4-bit bitsandbytes build): pip install transformers peft accelerate bitsandbytes.
python
import torchfrom peft import AutoPeftModelForCausalLMfrom transformers import AutoTokenizer repo = "ManideepVarma003/finexpert-qwen3-4b-lora"model = AutoPeftModelForCausalLM.from_pretrained(repo, device_map="auto")tokenizer = AutoTokenizer.from_pretrained(repo) SYSTEM = ("You are FinExpert, a specialized financial analysis assistant. Use only the financial information " "provided in the user input. Show relevant quantitative calculations when appropriate, distinguish " "reported facts from interpretation, and avoid inventing financial information or unsupported causal claims.")INSTRUCTION = ('Answer the question using only the context above. Show your working first, then give the final ' 'answer on the last line in the form "Answer: <value>".') context = "| ($ million) | 2023 | 2022 |\n|---|---|---|\n| Revenue | 1,250 | 1,100 |"question = "What was the percentage change in revenue from 2022 to 2023?" # Same ChatML layout as training (built by hand, no <think> block)prompt = (f"<|im_start|>system\n{SYSTEM}<|im_end|>\n" f"<|im_start|>user\n{context}\n\nQuestion: {question}\n\n{INSTRUCTION}<|im_end|>\n" f"<|im_start|>assistant\n")inputs = tokenizer(prompt, return_tensors="pt").to(model.device)with torch.no_grad(): out = model.generate(**inputs, max_new_tokens=384, do_sample=False)print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))Not for: investment advice, figures not present in the prompt, or any decision without human review. The model can miscalculate (see Limitations).
Give the context, then the question, and ask for the working followed by a final line Answer: <value>; the training prompt format is reproduced exactly by finexpert.evaluation.prompting in the repository. Use greedy decoding.
| Item | Value |
|---|---|
| Base | unsloth/Qwen3-4B-Instruct-2507-unsloth-bnb-4bit (4-bit) |
| Method | QLoRA, r 16, alpha 16, dropout 0, all attention and MLP projections; loss on the answer only |
| Data | 300 FinQA + 500 TAT-QA training questions (audited; FinQA gold = executed program) + 240 synthetic examples |
| Recipe | 1 epoch, LR 2e-4, effective batch 8, linear schedule, warmup 5%, AdamW 8-bit, max length 2,048 |
| Hardware | 1× Tesla T4 (Kaggle), about 35 minutes |
Validation subset (300 FinQA + 300 TAT-QA, never trained on):
| Model | FinQA accuracy | TAT-QA EM / F1 |
|---|---|---|
| Base model + answer-format rules (no training) | 63.3% | 61.3% / 67.2% |
| This adapter (v2-A, seed 42) | 60.0% | 66.0% / 73.8% |
Official test sets (scored once for this adapter):
| Test set | Result |
|---|---|
| FinQA (1,147 questions) | 62.0% answer accuracy (95% CI 59.1–64.8) |
| TAT-QA (1,663 questions, official scorer) | 62.1% EM / 70.2% F1 |
For reference on TAT-QA: TAGOP (the dataset paper's model) 50.1 / 58.0; GPT-4 zero-shot 71.9 / 79.7; human experts 84.1 / 90.8.
Selected among four fine-tuned runs by a rule registered before the final run (see the repository's experiment log). On these QA benchmarks the adapter matches a well-prompted base model (no significant difference) while answering in about 1/6 of the tokens; it also produces the structured synthetic-task outputs the base model cannot (classification 6–8/10 vs 0/10).
import torchfrom peft import AutoPeftModelForCausalLMfrom transformers import AutoTokenizer repo = "ManideepVarma003/finexpert-qwen3-4b-lora"model = AutoPeftModelForCausalLM.from_pretrained(repo, device_map="auto")tokenizer = AutoTokenizer.from_pretrained(repo) SYSTEM = ("You are FinExpert, a specialized financial analysis assistant. Use only the financial information " "provided in the user input. Show relevant quantitative calculations when appropriate, distinguish " "reported facts from interpretation, and avoid inventing financial information or unsupported causal claims.")INSTRUCTION = ('Answer the question using only the context above. Show your working first, then give the final ' 'answer on the last line in the form "Answer: <value>".') context = "| ($ million) | 2023 | 2022 |\n|---|---|---|\n| Revenue | 1,250 | 1,100 |"question = "What was the percentage change in revenue from 2022 to 2023?" # Same ChatML layout as training (built by hand, no <think> block)prompt = (f"<|im_start|>system\n{SYSTEM}<|im_end|>\n" f"<|im_start|>user\n{context}\n\nQuestion: {question}\n\n{INSTRUCTION}<|im_end|>\n" f"<|im_start|>assistant\n")inputs = tokenizer(prompt, return_tensors="pt").to(model.device)with torch.no_grad(): out = model.generate(**inputs, max_new_tokens=384, do_sample=False)print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))