owusu/DS-LLM-TEMPLATE-FINETUNING
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

instruct fine tuning setup

Browse Source
This commit is contained in:
Your Name
2025-08-28 22:41:56 +00:00
parent d49b4ff2d5
commit 78d7063ce5
58 changed files with 2510234 additions and 98 deletions
Download Patch File Download Diff File Expand all files Collapse all files
pipelines/instruct/.ipynb_checkpoints/train-checkpoint.py
+9 -9
View File
@@ -69,7 +69,7 @@ class InstructTrainer:
self.model_output_dir = config.get('model_output_dir', './models/instruct')
# Chat template
self.chat_template = config.get('chat_template', 'qwen-2.5')
self.chat_template = config.get('chat_template', 'llama-3.1')
def load_model_and_tokenizer(self):
"""Load the pre-trained model and tokenizer"""
@@ -240,13 +240,13 @@ class InstructTrainer:
try:
# For Qwen models, we need to use the correct chat template tokens
# Qwen uses different tokens than Llama
if "qwen" in self.model_name.lower():
instruction_part = "<|im_start|>user\n"
response_part = "<|im_start|>assistant\n"
else:
# Default for other models
instruction_part = "<|start_header_id|>user<|end_header_id|>\n\n"
response_part = "<|start_header_id|>assistant<|end_header_id|>\n\n"
# if "qwen" in self.model_name.lower():
instruction_part = "<|im_start|>user\n"
response_part = "<|im_start|>assistant\n"
# else:
# # Default for other models
# instruction_part = "<|start_header_id|>user<|end_header_id|>\n\n"
# response_part = "<|start_header_id|>assistant<|end_header_id|>\n\n"
# Configure trainer to only train on responses
self.trainer = train_on_responses_only(
@@ -291,7 +291,7 @@ class InstructTrainer:
print("Step 6: Setting up trainer...")
self.setup_trainer(formatted_dataset)
# Setup response-only training (optional but recommended for chat models)
#Setup response-only training (optional but recommended for chat models)
print("Step 7: Setting up response-only training...")
try:
self.setup_response_only_training()