Repository metrics
- Stars
- (5,126 個のスター)
- PR merge metrics
- (平均マージ 12h 49m) (30d で 2 merged PRs)
説明
使用纯GRPO进行训练，val中的acc和模型merge以后推理acc基本相同。 使用SFT后的ckp再进行GRPO，val中的acc很高，但训练后merge再推理，发现有严重的复读现象，且acc相差很大。
推理代码如下：
from transformers import Qwen2_5_VLForConditionalGeneration, AutoTokenizer, AutoProcessor from qwen_vl_utils import process_vision_info import re import pandas as pd import base64 from tqdm import tqdm
df = pd.read_parquet('val.parquet')
model_dir="SFT_GRPO/global90"
model = Qwen2_5_VLForConditionalGeneration.from_pretrained( model_dir, torch_dtype="auto", device_map="auto")
processor = AutoProcessor.from_pretrained(model_dir)
results = []
for inx in tqdm(range(len(df))):
img_bytes = (df.iloc[inx].images)[0]['bytes']
img_base64 = base64.b64encode(img_bytes).decode('utf-8')
data_url = f"data:image/png;base64,{img_base64}"
messages = [
{
"role": "system",
"content": "You are a helpful assistant."
},
{
"role": "user",
"content": [
{
"type": "image",
"image": data_url,
},
{"type": "text", "text": "my prompt"},
],
}
]
text = processor.apply_chat_template(
messages, tokenize=False, add_generation_prompt=True
)
image_inputs, video_inputs = process_vision_info(messages)
inputs = processor(
text=[text],
images=image_inputs,
videos=video_inputs,
padding=True,
return_tensors="pt",
)
inputs = inputs.to("cuda")
generated_ids = model.generate(**inputs, max_new_tokens=1024)
generated_ids_trimmed = [
out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output_text = processor.batch_decode(
generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
)
match = re.search(r'\boxed{([^}]+)}', output_text[0])
pred = match.group(1) if match else "ERROR"
actual = df.iloc[inx].answer
results.append(f"Predicted: {pred}, Actual: {actual}\n")
print((f"index: {inx} Predicted: {pred}, Actual: {actual}\n"), flush=True)