How sarandharshansp/amazon-ml-challenge-2025 works
This repository is a collection of experiments and submissions for the Amazon ML Challenge 2025 Smart Product Pricing Challenge. The principal workflow combines product text and images using CLIP and DistilBERT, trains a contrastive multimodal price regressor, produces predictions, and averages the two best models for the final submission. The README also documents CLIP-only, BERT-only, embedding-based, and feature-engineered approaches. No source excerpts were available, so links follow documented model designs and named notebook artifacts rather than inspected runtime calls.
Main entry points: amritha.ipynb CLIP Features · Averaging_code.ipynb Best-Model Average · predictions_amritha.csv Model Predictions · test_predictions_clip_big_epoch-10_full_v2_saran.csv CLIP Predictions
Components 15
Winning Multimodal Model
- CLIP Features
13 Oct 25 0859 PM IST/amritha.ipynb - DistilBERT Features
13 Oct 25 0859 PM IST/amritha.ipynb - Contrastive Learning
13 Oct 25 0859 PM IST/amritha.ipynb - Feature Fusion
13 Oct 25 0859 PM IST/amritha.ipynb - Price Regression
13 Oct 25 0859 PM IST/amritha.ipynb
Alternative Models
- CLIP Regression
13 Oct 25 0827 AM IST/clip-big-v2.ipynb - BERT Regression
12 Oct 25 0902 PM IST/large-bert.ipynb - DistilBERT Regression
12 Oct 25 0820 PM IST/distilbert-base-uncased-finetuned-sst-2-english.ipynb - Embedding Regression
11 Oct 25 0506 PM IST/qwen_06B_raw/Qwen3-0.6B.ipynb - Engineered Features
11 Oct 25 1129 PM IST/Untitled (6).ipynb
Prediction Artifacts
- Model Predictions
13 Oct 25 1139 PM IST - Best/predictions_amritha.csv - CLIP Predictions
13 Oct 25 1139 PM IST - Best/test_predictions_clip_big_epoch-10_full_v2_saran.csv - Best-Model Average
13 Oct 25 1139 PM IST - Best/Averaging_code.ipynb - Final Predictions
13 Oct 25 1139 PM IST - Best/final_prediction.csv
Other components
- Product Text + Images
Connections 16
- Product Text + Images → CLIP Features: encodes text and images
- Product Text + Images → DistilBERT Features: encodes text views
- CLIP Features → Contrastive Learning: aligns image-text features
- DistilBERT Features → Contrastive Learning: aligns augmented views
- CLIP Features → Feature Fusion: provides features
- DistilBERT Features → Feature Fusion: provides projection
- Feature Fusion → Price Regression: predicts price
- Price Regression → Model Predictions: writes predictions
- Product Text + Images → CLIP Regression: processes text and images
- Product Text + Images → BERT Regression: processes text
- Product Text + Images → DistilBERT Regression: processes text
- Product Text + Images → Embedding Regression: embeds text
- Product Text + Images → Engineered Features: extracts text features
- Model Predictions → Best-Model Average: averages model output
- CLIP Predictions → Best-Model Average: averages model output
- Best-Model Average → Final Predictions: writes final output