Ed Donner
GitHub
@@ -0,0 +1,262 @@
|
||||
# -*- coding: utf-8 -*-
|
||||
"""new_training_with_RAG.ipynb
|
||||
|
||||
Automatically generated by Colab.
|
||||
|
||||
Original file is located at
|
||||
https://colab.research.google.com/drive/1gi8FPI1dtnxBNTf86JdmXQ0BYqnKz7LS
|
||||
|
||||
# Predict Product Prices
|
||||
"""
|
||||
|
||||
!nvidia-smi
|
||||
|
||||
!pip install -q datasets requests torch peft bitsandbytes transformers trl accelerate sentencepiece matplotlib langchain-community chromadb
|
||||
|
||||
import os
|
||||
import re
|
||||
import math
|
||||
from tqdm import tqdm
|
||||
from google.colab import userdata
|
||||
from huggingface_hub import login
|
||||
import torch
|
||||
import transformers
|
||||
from transformers import (
|
||||
AutoModelForCausalLM, AutoTokenizer, TrainingArguments,
|
||||
set_seed, BitsAndBytesConfig, GenerationConfig)
|
||||
|
||||
from datasets import load_dataset
|
||||
from peft import LoraConfig, PeftModel
|
||||
from trl import SFTTrainer, SFTConfig
|
||||
from datetime import datetime
|
||||
import matplotlib.pyplot as plt
|
||||
|
||||
#LangChain & RAG Imports
|
||||
|
||||
from sentence_transformers import SentenceTransformer
|
||||
from langchain.schema import Document
|
||||
from langchain.vectorstores import Chroma
|
||||
import chromadb
|
||||
from langchain.embeddings import HuggingFaceEmbeddings
|
||||
|
||||
# Commented out IPython magic to ensure Python compatibility.
|
||||
# Constants
|
||||
|
||||
BASE_MODEL = "meta-llama/Meta-Llama-3.1-8B"
|
||||
#BASE_MODEL = 'mistralai/Mistral-7B-Instruct-v0.1'
|
||||
PROJECT_NAME = "pricer-optim"
|
||||
HF_USER = "Adriana213"
|
||||
|
||||
# Data
|
||||
|
||||
DATASET_NAME = f"{HF_USER}/pricer-data"
|
||||
MAX_SEQUENCE_LENGTH = 182
|
||||
|
||||
|
||||
RUN_NAME = f"{PROJECT_NAME}-{datetime.now():%Y%m%d_%H%M%S}"
|
||||
|
||||
HUB_MODEL_NAME = f"{HF_USER}/{RUN_NAME}"
|
||||
|
||||
# Hyperparameters for QLoRA
|
||||
|
||||
LORA_R = 8
|
||||
LORA_ALPHA = 32
|
||||
TARGET_MODULES = ["q_proj", "v_proj", "k_proj", "o_proj"]
|
||||
LORA_DROPOUT = 0.10
|
||||
QUANT_4_BIT = True
|
||||
|
||||
# Hyperparameters for Training
|
||||
|
||||
EPOCHS = 2
|
||||
BATCH_SIZE = 16
|
||||
GRADIENT_ACCUMULATION_STEPS = 1
|
||||
LEARNING_RATE = 2e-4
|
||||
LR_SCHEDULER_TYPE = 'cosine'
|
||||
WARMUP_RATIO = 0.05
|
||||
OPTIMIZER = "paged_adamw_32bit"
|
||||
STEPS = 50
|
||||
SAVE_STEPS = 200
|
||||
EVAL_STEPS = 200 # kept for potential future use
|
||||
|
||||
# %matplotlib inline
|
||||
|
||||
HUB_MODEL_NAME
|
||||
|
||||
"""### Log in to HuggingFace & get Data"""
|
||||
|
||||
hf_token = userdata.get('HF_TOKEN')
|
||||
login(hf_token, add_to_git_credential=True)
|
||||
|
||||
torch.cuda.empty_cache()
|
||||
|
||||
dataset = load_dataset(DATASET_NAME)
|
||||
train = dataset['train']
|
||||
test = dataset['test']
|
||||
|
||||
"""## Now load the Tokenizer and Model
|
||||
|
||||
The model is "quantized" - we are reducing the precision to 4 bits.
|
||||
"""
|
||||
|
||||
# Pick the right quantization
|
||||
|
||||
if QUANT_4_BIT:
|
||||
quant_config = BitsAndBytesConfig(
|
||||
load_in_4bit=True,
|
||||
bnb_4bit_use_double_quant=True,
|
||||
bnb_4bit_compute_dtype=torch.bfloat16,
|
||||
bnb_4bit_quant_type="nf4"
|
||||
)
|
||||
else:
|
||||
quant_config = BitsAndBytesConfig(
|
||||
load_in_8bit=True,
|
||||
bnb_8bit_compute_dtype=torch.bfloat16
|
||||
)
|
||||
|
||||
# Load the Tokenizer and the Model
|
||||
|
||||
tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL, trust_remote_code=True)
|
||||
tokenizer.pad_token = tokenizer.eos_token
|
||||
tokenizer.padding_side = "right"
|
||||
|
||||
base_model = AutoModelForCausalLM.from_pretrained(
|
||||
BASE_MODEL,
|
||||
quantization_config=quant_config,
|
||||
device_map="auto",
|
||||
)
|
||||
|
||||
base_model.generation_config.pad_token_id = tokenizer.pad_token_id
|
||||
|
||||
print(f"Memory footprint: {base_model.get_memory_footprint() / 1e6:.1f} MB")
|
||||
|
||||
"""# Data Collator
|
||||
|
||||
"""
|
||||
|
||||
from trl import DataCollatorForCompletionOnlyLM
|
||||
|
||||
response_template = "Price is $"
|
||||
collator = DataCollatorForCompletionOnlyLM(response_template,
|
||||
tokenizer=tokenizer)
|
||||
|
||||
"""# Set up the configuration for Training"""
|
||||
|
||||
# LoRA Config
|
||||
|
||||
lora_parameters = LoraConfig(
|
||||
lora_alpha = LORA_ALPHA,
|
||||
lora_dropout = LORA_DROPOUT,
|
||||
r = LORA_R,
|
||||
bias = "none",
|
||||
task_type = "CAUSAL_LM",
|
||||
target_modules = TARGET_MODULES,
|
||||
)
|
||||
|
||||
# Training Config
|
||||
|
||||
train_parameters = SFTConfig(
|
||||
output_dir = RUN_NAME,
|
||||
num_train_epochs = EPOCHS,
|
||||
per_device_train_batch_size = BATCH_SIZE,
|
||||
per_device_eval_batch_size = 4,
|
||||
eval_strategy = "no",
|
||||
eval_steps = EVAL_STEPS,
|
||||
gradient_accumulation_steps = GRADIENT_ACCUMULATION_STEPS,
|
||||
optim = OPTIMIZER,
|
||||
save_steps = SAVE_STEPS,
|
||||
save_total_limit = 5,
|
||||
logging_steps = 50,
|
||||
learning_rate = LEARNING_RATE,
|
||||
weight_decay = 0.01,
|
||||
fp16=False,
|
||||
bf16=True,
|
||||
max_grad_norm=0.3,
|
||||
max_steps=-1,
|
||||
warmup_ratio = WARMUP_RATIO,
|
||||
group_by_length=True,
|
||||
lr_scheduler_type = LR_SCHEDULER_TYPE,
|
||||
run_name = RUN_NAME,
|
||||
max_seq_length = MAX_SEQUENCE_LENGTH,
|
||||
dataset_text_field = "text",
|
||||
save_strategy = "steps",
|
||||
hub_strategy = "every_save",
|
||||
push_to_hub = True,
|
||||
hub_model_id = HUB_MODEL_NAME,
|
||||
hub_private_repo = True,
|
||||
report_to = 'none',
|
||||
)
|
||||
|
||||
|
||||
fine_tuning = SFTTrainer(
|
||||
model = base_model,
|
||||
train_dataset = train,
|
||||
eval_dataset=test,
|
||||
peft_config = lora_parameters,
|
||||
args = train_parameters,
|
||||
data_collator = collator,
|
||||
)
|
||||
|
||||
"""## Fine Tuning"""
|
||||
|
||||
fine_tuning.train()
|
||||
|
||||
fine_tuning.model.push_to_hub(RUN_NAME, private=True)
|
||||
print(f"Saved to the hub: {RUN_NAME}")
|
||||
|
||||
"""# Implement RAG"""
|
||||
|
||||
HF_USER = "Adriana213"
|
||||
RUN_NAME = "pricer-optim-20250514_061529"
|
||||
fine_tuned_model = PeftModel.from_pretrained(base_model, f"{HF_USER}/{RUN_NAME}")
|
||||
print(f"✅ Loaded fine-tuned adapter: {HF_USER}/{RUN_NAME}")
|
||||
|
||||
base_model = fine_tuned_model
|
||||
|
||||
"""## Build Chroma index"""
|
||||
|
||||
docs = [
|
||||
Document(page_content=text, metadata = {'price': price})
|
||||
for text, price in zip(train['text'], train['price'])
|
||||
]
|
||||
|
||||
# Create embeddings & persist Chroma index
|
||||
|
||||
embedding = HuggingFaceEmbeddings(model_name = 'all-MiniLM-L6-v2')
|
||||
chroma = Chroma.from_documents(
|
||||
documents = docs,
|
||||
embedding = embedding,
|
||||
persist_directory = 'chroma_train_index'
|
||||
)
|
||||
|
||||
chroma.persist()
|
||||
print('Chroma index built and persisted.')
|
||||
|
||||
"""## RAG Prediction Function"""
|
||||
|
||||
generation_config = GenerationConfig(
|
||||
max_new_token = 10,
|
||||
do_sample = False,
|
||||
temperature = 0.1
|
||||
)
|
||||
|
||||
def predict_price_rag(desc: str, k: int = 3) -> float:
|
||||
hits = chroma.similarity_search(desc, k = k)
|
||||
shot_strs = [
|
||||
f'Description: {doc.page_content}\nPrice is ${doc.metadata["price"]}'
|
||||
for doc in hits
|
||||
]
|
||||
|
||||
prompt = "\n\n".join(shot_strs) + f"\n\nDescription: {desc}\nPrice is $"
|
||||
|
||||
inputs = tokenizer(prompt, return_tensors="pt").to(base_model.device)
|
||||
out = base_model.generate(**inputs, generation_config=generation_config)
|
||||
text = tokenizer.decode(
|
||||
out[0, inputs["input_ids"].shape[-1]:],
|
||||
skip_special_tokens=True
|
||||
).strip()
|
||||
return float(re.findall(r"\d+\.?\d+", text)[0])
|
||||
|
||||
!zip -r chroma_index.zip chroma_train_index
|
||||
|
||||
from google.colab import files
|
||||
files.download("chroma_index.zip")
|
||||
@@ -0,0 +1,258 @@
|
||||
# -*- coding: utf-8 -*-
|
||||
"""Testing Fine-tuned model with RAG
|
||||
|
||||
Automatically generated by Colab.
|
||||
|
||||
Original file is located at
|
||||
https://colab.research.google.com/drive/1J8P8cwqwhBo3CNIZaEFe6BMRw0WUfEqy
|
||||
|
||||
## Predict Product Prices
|
||||
|
||||
### And now, to evaluate our fine-tuned open source model
|
||||
"""
|
||||
|
||||
!pip install -q datasets peft requests torch bitsandbytes transformers trl accelerate sentencepiece matplotlib langchain-community chromadb
|
||||
|
||||
import os
|
||||
import re
|
||||
import math
|
||||
|
||||
from google.colab import userdata
|
||||
|
||||
from huggingface_hub import login
|
||||
|
||||
import torch
|
||||
import torch.nn.functional as F
|
||||
|
||||
from transformers import (
|
||||
AutoModelForCausalLM, AutoTokenizer,
|
||||
BitsAndBytesConfig, GenerationConfig)
|
||||
|
||||
from datasets import load_dataset
|
||||
|
||||
from peft import PeftModel
|
||||
|
||||
from sentence_transformers import SentenceTransformer
|
||||
from langchain.vectorstores import Chroma
|
||||
from langchain.embeddings import HuggingFaceEmbeddings
|
||||
|
||||
import matplotlib.pyplot as plt
|
||||
|
||||
# Commented out IPython magic to ensure Python compatibility.
|
||||
# Constants
|
||||
|
||||
BASE_MODEL = "meta-llama/Llama-3.1-8B"
|
||||
PROJECT_NAME = "pricer"
|
||||
HF_USER = "Adriana213"
|
||||
|
||||
RUN_NAME = "optim-20250514_061529"
|
||||
PROJECT_RUN_NAME = f"{PROJECT_NAME}-{RUN_NAME}"
|
||||
|
||||
FINETUNED_MODEL = f"{HF_USER}/{PROJECT_RUN_NAME}"
|
||||
|
||||
# Data
|
||||
|
||||
DATASET_NAME = f"{HF_USER}/pricer-data"
|
||||
|
||||
# Hyperparameters for QLoRA
|
||||
|
||||
QUANT_4_BIT = True
|
||||
|
||||
# %matplotlib inline
|
||||
|
||||
# Used for writing to output in color
|
||||
|
||||
GREEN = "\033[92m"
|
||||
YELLOW = "\033[93m"
|
||||
RED = "\033[91m"
|
||||
RESET = "\033[0m"
|
||||
COLOR_MAP = {"red":RED, "orange": YELLOW, "green": GREEN}
|
||||
|
||||
"""### Log in to HuggingFace
|
||||
|
||||
|
||||
"""
|
||||
|
||||
hf_token = userdata.get('HF_TOKEN')
|
||||
login(hf_token, add_to_git_credential=True)
|
||||
|
||||
dataset = load_dataset(DATASET_NAME)
|
||||
train = dataset['train']
|
||||
test = dataset['test']
|
||||
|
||||
test[0]
|
||||
|
||||
"""## Now load the Tokenizer and Model"""
|
||||
|
||||
if QUANT_4_BIT:
|
||||
quant_config = BitsAndBytesConfig(
|
||||
load_in_4bit=True,
|
||||
bnb_4bit_use_double_quant=True,
|
||||
bnb_4bit_compute_dtype=torch.bfloat16,
|
||||
bnb_4bit_quant_type="nf4"
|
||||
)
|
||||
else:
|
||||
quant_config = BitsAndBytesConfig(
|
||||
load_in_8bit=True,
|
||||
bnb_8bit_compute_dtype=torch.bfloat16
|
||||
)
|
||||
|
||||
# Load the Tokenizer and the Model
|
||||
|
||||
tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL, trust_remote_code=True)
|
||||
tokenizer.pad_token = tokenizer.eos_token
|
||||
tokenizer.padding_side = "right"
|
||||
|
||||
base_model = AutoModelForCausalLM.from_pretrained(
|
||||
BASE_MODEL,
|
||||
quantization_config=quant_config,
|
||||
device_map="auto",
|
||||
)
|
||||
base_model.generation_config.pad_token_id = tokenizer.pad_token_id
|
||||
|
||||
# Load the fine-tuned model with PEFT
|
||||
|
||||
fine_tuned_model = PeftModel.from_pretrained(base_model, FINETUNED_MODEL)
|
||||
|
||||
|
||||
print(f"Memory footprint: {fine_tuned_model.get_memory_footprint() / 1e6:.1f} MB")
|
||||
|
||||
fine_tuned_model
|
||||
|
||||
"""# Evaluation"""
|
||||
|
||||
def extract_price(s):
|
||||
if "Price is $" in s:
|
||||
contents = s.split("Price is $")[1]
|
||||
contents = contents.replace(',','')
|
||||
match = re.search(r"[-+]?\d*\.\d+|\d+", contents)
|
||||
return float(match.group()) if match else 0
|
||||
return 0
|
||||
|
||||
extract_price("Price is $a fabulous 899.99 or so")
|
||||
|
||||
# Original prediction function takes the most likely next token
|
||||
|
||||
def model_predict(prompt):
|
||||
inputs = tokenizer.encode(prompt, return_tensors="pt").to("cuda")
|
||||
attention_mask = torch.ones(inputs.shape, device="cuda")
|
||||
outputs = fine_tuned_model.generate(inputs, attention_mask=attention_mask, max_new_tokens=3, num_return_sequences=1)
|
||||
response = tokenizer.decode(outputs[0])
|
||||
return extract_price(response)
|
||||
|
||||
# top_K = 3
|
||||
|
||||
# def improved_model_predict(prompt, device="cuda"):
|
||||
# set_seed(42)
|
||||
# inputs = tokenizer.encode(prompt, return_tensors="pt").to(device)
|
||||
# attention_mask = torch.ones(inputs.shape, device=device)
|
||||
|
||||
# with torch.no_grad():
|
||||
# outputs = fine_tuned_model(inputs, attention_mask=attention_mask)
|
||||
# next_token_logits = outputs.logits[:, -1, :].to('cpu')
|
||||
|
||||
# next_token_probs = F.softmax(next_token_logits, dim=-1)
|
||||
# top_prob, top_token_id = next_token_probs.topk(top_K)
|
||||
# prices, weights = [], []
|
||||
# for i in range(top_K):
|
||||
# predicted_token = tokenizer.decode(top_token_id[0][i])
|
||||
# probability = top_prob[0][i]
|
||||
# try:
|
||||
# result = float(predicted_token)
|
||||
# except ValueError as e:
|
||||
# result = 0.0
|
||||
# if result > 0:
|
||||
# prices.append(result)
|
||||
# weights.append(probability)
|
||||
# if not prices:
|
||||
# return 0.0, 0.0
|
||||
# total = sum(weights)
|
||||
# weighted_prices = [price * weight / total for price, weight in zip(prices, weights)]
|
||||
# return sum(weighted_prices).item()
|
||||
|
||||
embedder = HuggingFaceEmbeddings(model_name = "all-MiniLM-L6-v2")
|
||||
chroma = Chroma(
|
||||
persist_directory = "chroma_train_index",
|
||||
embedding_function = embedder
|
||||
)
|
||||
|
||||
gen_config = GenerationConfig(max_new_tokens=10, do_sample=False)
|
||||
|
||||
def predict_price_rag(desc: str, k: int = 3) -> float:
|
||||
docs = chroma.similarity_search(desc, k=k)
|
||||
shots = "\n\n".join(f"Description: {d.page_content}\nPrice is ${d.metadata['price']}"
|
||||
for d in docs)
|
||||
prompt = f"{shots}\n\nDescription: {desc}\nPrice is $"
|
||||
inp = tokenizer(prompt, return_tensors="pt").to(fine_tuned_model.device)
|
||||
out = fine_tuned_model.generate(**inp, generation_config=gen_config)
|
||||
txt = tokenizer.decode(out[0, inp["input_ids"].shape[-1]:], skip_special_tokens=True).strip()
|
||||
return float(re.findall(r"\d+\.?\d+", txt)[0])
|
||||
|
||||
class Tester:
|
||||
|
||||
def __init__(self, predictor, data, title=None, size=250):
|
||||
self.predictor = predictor
|
||||
self.data = data
|
||||
self.title = title or predictor.__name__.replace("_", " ").title()
|
||||
self.size = size
|
||||
self.guesses = []
|
||||
self.truths = []
|
||||
self.errors = []
|
||||
self.sles = []
|
||||
self.colors = []
|
||||
|
||||
def color_for(self, error, truth):
|
||||
if error<40 or error/truth < 0.2:
|
||||
return "green"
|
||||
elif error<80 or error/truth < 0.4:
|
||||
return "orange"
|
||||
else:
|
||||
return "red"
|
||||
|
||||
def run_datapoint(self, i):
|
||||
datapoint = self.data[i]
|
||||
guess = self.predictor(datapoint["text"])
|
||||
truth = datapoint["price"]
|
||||
error = abs(guess - truth)
|
||||
log_error = math.log(truth+1) - math.log(guess+1)
|
||||
sle = log_error ** 2
|
||||
color = self.color_for(error, truth)
|
||||
title = datapoint["text"].split("\n\n")[1][:20] + "..."
|
||||
self.guesses.append(guess)
|
||||
self.truths.append(truth)
|
||||
self.errors.append(error)
|
||||
self.sles.append(sle)
|
||||
self.colors.append(color)
|
||||
print(f"{COLOR_MAP[color]}{i+1}: Guess: ${guess:,.2f} Truth: ${truth:,.2f} Error: ${error:,.2f} SLE: {sle:,.2f} Item: {title}{RESET}")
|
||||
|
||||
def chart(self, title):
|
||||
max_error = max(self.errors)
|
||||
plt.figure(figsize=(12, 8))
|
||||
max_val = max(max(self.truths), max(self.guesses))
|
||||
plt.plot([0, max_val], [0, max_val], color='deepskyblue', lw=2, alpha=0.6)
|
||||
plt.scatter(self.truths, self.guesses, s=3, c=self.colors)
|
||||
plt.xlabel('Ground Truth')
|
||||
plt.ylabel('Model Estimate')
|
||||
plt.xlim(0, max_val)
|
||||
plt.ylim(0, max_val)
|
||||
plt.title(title)
|
||||
plt.show()
|
||||
|
||||
def report(self):
|
||||
average_error = sum(self.errors) / self.size
|
||||
rmsle = math.sqrt(sum(self.sles) / self.size)
|
||||
hits = sum(1 for color in self.colors if color=="green")
|
||||
title = f"{self.title} Error=${average_error:,.2f} RMSLE={rmsle:,.2f} Hits={hits/self.size*100:.1f}%"
|
||||
self.chart(title)
|
||||
|
||||
def run(self):
|
||||
self.error = 0
|
||||
for i in range(self.size):
|
||||
self.run_datapoint(i)
|
||||
self.report()
|
||||
|
||||
@classmethod
|
||||
def test(cls, function, data):
|
||||
cls(function, data).run()
|
||||
|
||||
Tester.test(predict_price_rag, test)
|
||||
Reference in New Issue
Block a user