🧠 两小时从 0 训出你的 64M 大模型。Train your own 64M LLM from scratch in 2 hours.
[update] align mask
jingyaogongaa539a824add
diff --git a/dataset/lm_dataset.py b/dataset/lm_dataset.pyindex 21e321c..51db696 100644--- a/dataset/lm_dataset.py+++ b/dataset/lm_dataset.py@@ -35,8 +35,8 @@ class SFTDataset(Dataset):self.tokenizer = tokenizerself.max_length = max_lengthself.samples = load_dataset('json', data_files=jsonl_path, split='train')- self.bos_id = tokenizer(f'{tokenizer.bos_token}assistant', add_special_tokens=False).input_ids- self.eos_id = tokenizer(f'{tokenizer.eos_token}', add_special_tokens=False).input_ids+ self.bos_id = tokenizer(f'{tokenizer.bos_token}assistant\n', add_special_tokens=False).input_ids+ self.eos_id = tokenizer(f'{tokenizer.eos_token}\n', add_special_tokens=False).input_idsdef __len__(self):return len(self.samples)@@ -62,7 +62,7 @@ class SFTDataset(Dataset):if input_ids[end:end + len(self.eos_id)] == self.eos_id:breakend += 1- for j in range(start + 1, min(end + len(self.eos_id) + 1, self.max_length)):+ for j in range(start, min(end + len(self.eos_id), self.max_length)):labels[j] = input_ids[j]i = end + len(self.eos_id) if end < len(input_ids) else len(input_ids)else:@@ -89,8 +89,8 @@ class DPODataset(Dataset):self.tokenizer = tokenizerself.max_length = max_lengthself.padding = tokenizer.pad_token_id if tokenizer.pad_token_id is not None else 0- self.bos_id = tokenizer(f'{tokenizer.bos_token}assistant', add_special_tokens=False).input_ids- self.eos_id = tokenizer(f'{tokenizer.eos_token}', add_special_tokens=False).input_ids+ self.bos_id = tokenizer(f'{tokenizer.bos_token}assistant\n', add_special_tokens=False).input_ids+ self.eos_id = tokenizer(f'{tokenizer.eos_token}\n', add_special_tokens=False).input_idsself.data = load_dataset('json', data_files=file_path, split='train')def __len__(self):@@ -146,7 +146,7 @@ class DPODataset(Dataset):if input_ids[end:end + len(self.eos_id)] == self.eos_id:breakend += 1- for j in range(start + 1, min(end + len(self.eos_id) + 1, self.max_length)):+ for j in range(start, min(end + len(self.eos_id), self.max_length)):loss_mask[j] = 1i = end + len(self.eos_id) if end < len(input_ids) else len(input_ids)else: