from transformers import AutoTokenizer, AutoModelForImageTextToText
import torch

model_id = "deadbear34/qwen35-4b-plantdisease-merged"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="cuda",
    trust_remote_code=True,
)

prompt = "<|user|>
Apa penyebab late blight pada tomat?
<|assistant|>
"
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
output = model.generate(**inputs, max_new_tokens=512, do_sample=False)
print(tokenizer.decode(output[0], skip_special_tokens=True))

Benchmark Results (STANDARD mode evaluation)

Base CPT: deadbear34/qwen35-4b-plantdisease-cpt
LoRA SFT: deadbear34/qwen35-4b-plantdisease-sft-lora
Tokenizer source: Qwen/Qwen3.5-4B-Base (original Qwen, vocab=248,044)
Merged at: 2026-05-13 14:32

Benchmark	Score	Notes
GSM8K (n=300)	90.00%	Math reasoning with thinking format
MMLU (12 subjects, n=1000)	76.00%	Biology/medical-skewed subset
HumanEval (full 164)	42.68%	Coding (regressed from base)
TruthfulQA-MC1 (n=300)	57.00%	Truthfulness
PlantDisease ROUGE-1 (n=100)	38.39%	Domain-specific, EN+ID

⚠️ Note: MMLU evaluated on biology-heavy subset (12/57 subjects). Full MMLU likely lower.

License

Apache 2.0