Spaces:
Running
Running
File size: 4,532 Bytes
7b766ee |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 |
#!/usr/bin/env python3
import os
from torch.utils.data import DataLoader
from sentence_transformers import SentenceTransformer, InputExample, losses
# If you want to push to the HF Hub/Spaces programmatically:
# pip install huggingface_hub
# from huggingface_hub import HfApi, HfFolder
def main():
#--------------------------------------------------------------------------
# 1. (Optional) Setup your Hugging Face auth
#--------------------------------------------------------------------------
# If you need to log into your HF account, you can do:
# hf_token = os.getenv("HF_TOKEN") # or read from a config file
# HfFolder.save_token(hf_token)
# api = HfApi()
#
# Then set something like:
# repo_id = "KolumbusLindh/my-weekly-model"
#
# Alternatively, you can push manually later via huggingface-cli.
#--------------------------------------------------------------------------
# 2. Placeholder training data
#--------------------------------------------------------------------------
# Suppose each tuple is: (CV_text, liked_job_text, disliked_job_text).
# In a real scenario, you'd gather user feedback from your database.
train_data = [
("My CV #1", "Job #1 that user liked", "Job #1 that user disliked"),
("My CV #2", "Job #2 that user liked", "Job #2 that user disliked"),
# ...
]
#--------------------------------------------------------------------------
# 3. Convert data into Sentence Transformers InputExamples
#--------------------------------------------------------------------------
train_examples = []
for (cv_text, liked_job_text, disliked_job_text) in train_data:
example = InputExample(
texts=[cv_text, liked_job_text, disliked_job_text]
# TripletLoss expects exactly 3 texts: anchor, positive, negative
)
train_examples.append(example)
#--------------------------------------------------------------------------
# 4. Load the base model
#--------------------------------------------------------------------------
model_name = "sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2"
model = SentenceTransformer(model_name)
#--------------------------------------------------------------------------
# 5. Prepare DataLoader & define the Triplet Loss
#--------------------------------------------------------------------------
# A typical margin is 0.5–1.0. Feel free to adjust it.
train_dataloader = DataLoader(train_examples, shuffle=True, batch_size=8)
train_loss = losses.TripletLoss(
model=model,
distance_metric=losses.TripletDistanceMetric.COSINE,
margin=0.5
)
#--------------------------------------------------------------------------
# 6. Fine-tune (fit) the model
#--------------------------------------------------------------------------
# Just 1 epoch here for demo. In practice, tune #epochs/batch_size, etc.
num_epochs = 1
warmup_steps = int(len(train_dataloader) * num_epochs * 0.1) # ~10% warmup
model.fit(
train_objectives=[(train_dataloader, train_loss)],
epochs=num_epochs,
warmup_steps=warmup_steps,
show_progress_bar=True
)
#--------------------------------------------------------------------------
# 7. Save model locally
#--------------------------------------------------------------------------
local_output_path = "my_finetuned_model"
model.save(local_output_path)
print(f"Model fine-tuned and saved locally to: {local_output_path}")
#--------------------------------------------------------------------------
# 8. (Optional) Push to your Hugging Face Space
#--------------------------------------------------------------------------
# If you want to push automatically:
#
# model.push_to_hub(repo_id=repo_id, commit_message="Weekly model update")
#
# Or if you have a Space at e.g. https://huggingface.co/spaces/KolumbusLindh/<some-name>,
# you’d create a repo on HF, then push to that repo. Typically one uses
# huggingface-cli or the huggingface_hub methods for that:
#
# api.create_repo(repo_id=repo_id, repo_type="model", private=False)
# model.push_to_hub(repo_id=repo_id)
#
# # If it's a Space, you might need to store your model in the "models" folder
# # or however your Gradio app is set up to load it.
if __name__ == "__main__":
main()
|