ros/LLaMA-Factory
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

support lora for llama pro

Browse Source 
Former-commit-id: f74c78ba95f0545aae89e603e466f494705ad024
This commit is contained in:
hiyouga
2024-02-21 02:17:22 +08:00
parent a3f30038a0
commit bc16c9a54a
7 changed files with 119 additions and 28 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
src/llmtuner/data/preprocess.py
View File

@@ -99,12 +99,12 @@ def preprocess_packed_supervised_dataset(
continue
messages = examples["prompt"][i] + examples["response"][i]
for turn_idx, (source_ids, target_ids) in enumerate(
template.encode_multiturn(tokenizer, messages, examples["system"][i], examples["tools"][i])
for source_ids, target_ids in template.encode_multiturn(
tokenizer, messages, examples["system"][i], examples["tools"][i]
):
if data_args.train_on_prompt:
source_mask = source_ids
elif turn_idx != 0 and template.efficient_eos:
elif len(input_ids) != 0 and template.efficient_eos:
source_mask = [tokenizer.eos_token_id] + [IGNORE_INDEX] * (len(source_ids) - 1)
else:
source_mask = [IGNORE_INDEX] * len(source_ids)
@@ -112,9 +112,9 @@ def preprocess_packed_supervised_dataset(
input_ids += source_ids + target_ids
labels += source_mask + target_ids
if template.efficient_eos:
input_ids += [tokenizer.eos_token_id]
labels += [tokenizer.eos_token_id]
if template.efficient_eos:
input_ids += [tokenizer.eos_token_id]
labels += [tokenizer.eos_token_id]
total_length = len(input_ids)
block_size = data_args.cutoff_len
@@ -122,9 +122,10 @@ def preprocess_packed_supervised_dataset(
total_length = (total_length // block_size) * block_size
# split by chunks of cutoff_len
for i in range(0, total_length, block_size):
model_inputs["input_ids"].append(input_ids[i : i + block_size])
model_inputs["attention_mask"].append([1] * block_size)
model_inputs["labels"].append(labels[i : i + block_size])
if not all(label == IGNORE_INDEX for label in labels[i : i + block_size]):
model_inputs["input_ids"].append(input_ids[i : i + block_size])
model_inputs["attention_mask"].append([1] * block_size)
model_inputs["labels"].append(labels[i : i + block_size])
return model_inputs
@@ -180,7 +181,6 @@ def preprocess_pairwise_dataset(
chosen_messages = examples["prompt"][i] + [examples["response"][i][0]]
rejected_messages = examples["prompt"][i] + [examples["response"][i][1]]
prompt_ids, chosen_ids = template.encode_oneturn(
tokenizer,
chosen_messages,