###推論方法

model_id = "hara-CU/gemma9b-FT-Candidate241214-1"

dtype = None

load_in_4bit = True

model, tokenizer = FastLanguageModel.from_pretrained( model_name=model_id, dtype=dtype, load_in_4bit=load_in_4bit, trust_remote_code=True, ) FastLanguageModel.for_inference(model)

datasets ="file_path1" #データセットの読み込み

results = []

for dt in tqdm(datasets):

input = dt["input"]

prompt = f"""### 指示\n{input}\n### 回答\n"""

inputs = tokenizer([prompt], return_tensors = "pt").to(model.device)

outputs = model.generate(**inputs, max_new_tokens = 700, use_cache = True, do_sample=False, repetition_penalty=1.2) prediction = tokenizer.decode(outputs[0], skip_special_tokens=True).split('\n### 回答')[-1]

results.append({"task_id": dt["task_id"], "input": input, "output": prediction})

json_file="file_path2"

with open(json_file, 'w', encoding='utf-8') as f: for result in results: json.dump(result, f, ensure_ascii=False) f.write('\n')

##使用データセット

weblab-GENIAC/aya-ja-evol-instruct-calm3-dpo-masked

DeL-TaiseiOzaki/Tengentoppa-sft-reasoning-ja

This gemma2 model was trained 2x faster with Unsloth and Huggingface's TRL library.