Case studies / Fine-tuning

Fine-tuning

LoRA, distillation and training models on a company's own data.

Approach

17 write-ups

Pinterest6 Nov 2026Fine-tuning
Safe and Efficient Multi-Epoch Embedding Training
Problem Fine‑tuning embeddings for only one epoch limited model quality on fresh data.
Solution Created a multi‑epoch training pipeline that pre‑trains ID embeddings with a contrastive objective and then fine‑tunes them safely, avoiding memorization.
Read on medium.comDetails
NVIDIA27 Oct 2026NewFine-tuning
Fine-Tuning NVIDIA Nemotron to Know Local Rules, Not Just Local Languages
Problem The model gave generic, inaccurate answers for local French and Belgian queries, missing regional rules and cultural nuance.
Solution NVIDIA fine‑tuned Nemotron with country‑specific persona datasets, adding synthetic local demographics and regulations so the model answers with correct regional information.
Read on developer.nvidia.comDetails
Pinterest9 Oct 2026Fine-tuning
From Activity to Intent: Generating User Journeys with LLMs
Problem Pinterest needed to identify users' long‑term goals, not just the last pin they clicked.
Solution They fine‑tuned a 4 billion‑parameter LLM to read a user's activity timeline and output ranked journey names, distilling knowledge from a larger model and deploying it at notification scale.
Read on medium.comDetails
NVIDIA1 Oct 2026Fine-tuning
Fine-Tuning NVIDIA Nemotron for Saudi Arabic Dialects, with a Path to Other Languages
Problem ASR struggled with Saudi Arabic dialects and local recording conditions.
Solution They curated a small dialect corpus, mixed it with existing data, and fine‑tuned Nemotron 3.5 using NVIDIA NeMo’s efficient batching recipe.
Read on developer.nvidia.comDetails
Hugging Face3 Sep 2026Fine-tuning
Fine-tuning a 350M Model for Better Structured Outputs in 100 GRPO Steps
Problem Models often output data in the wrong format, preventing integration into downstream systems.
Solution Fine‑tuned a 350M LLM with Group Relative Policy Optimization using the TRL library, running only 100 steps on a free‑tier GPU, raising structured‑output accuracy on IFStruct.
Accuracy improved to 29.7%
Read on huggingface.coDetails
Google26 Aug 2026Fine-tuning
GlucoFM: Foundation model for continuous glucose monitoring
Problem Users lacked accurate ways to predict diabetes risk from cheap glucose data.
Solution They trained a lightweight self‑supervised model that splits glucose signals into slow trends and quick spikes, learning useful features from many unlabeled readings.
PR‑AUC improved 5.8 percentage points over the best prior model
Read on research.googleDetails
Hugging Face26 Aug 2026Fine-tuning
Training and Finetuning Multi-Vector Embedding Models with Sentence Transformers
Problem Search quality on medical texts was poor with generic models.
Solution Finetuned a multi‑vector encoder using Sentence Transformers, training it on domain data with custom loss and evaluation, achieving stronger token‑level matching.
Outperformed all general‑purpose retrievers
Research behind this, on AI Papers DailyReducing Representational Interference in Retrieval Embeddings with Trainable Prefix Tokens

Read on huggingface.coDetails
Shopify5 Aug 2026Fine-tuning
Sidekick's continual learning loop
Problem Frontier models were slow, costly, and couldn’t learn from real‑world mistakes.
Solution Created a continual learning loop that turns production failures into training data, fine‑tunes the model, and updates it automatically, cutting latency and cost.
Cost reduced 96%
Read on shopify.engineeringDetails
Dropbox20 Jul 2026Fine-tuning

How our universal content processing platform Riviera evolved for AI and beyond
Problem Users needed fast, consistent previews and AI-ready content across many file types.
Solution Built Riviera, a shared platform that converts files into thumbnails, PDFs, and lightweight videos, scaling to handle billions of transformations daily.

Read on dropbox.techDetails
JioHotstar1 Jul 2026Fine-tuning
Consistent Generative Video: Building Video Models That Remember
Problem Users wanted realistic talking‑head videos of IP characters without reshooting, but existing models drifted in identity and lip sync.
Solution They trained tiny LoRA adapters on two video diffusion models, keeping the base frozen and adding only ~0.1% new parameters per layer, enabling fast, scalable character‑specific generation.
Read on blog.hotstar.comDetails
Google30 Jun 2026Fine-tuning
Introducing TabFM: A zero-shot foundation model for tabular data
Problem Data scientists spent too much time training and tuning models for each new table.
Solution They built TabFM, a foundation model that uses in‑context learning to predict on any tabular data in one forward pass, removing the need for training or feature engineering.
Read on research.googleDetails
Hugging Face24 Jun 2026Fine-tuning
Accelerating Transformers Fine-Tuning with NVIDIA NeMo AutoModel
Problem Fine‑tuning large MoE models was slow and consumed too much GPU memory.
Solution Combined Hugging Face Transformers v5 with NVIDIA NeMo AutoModel, adding expert parallelism and fused kernels, to speed up training and cut memory using the same API.
3.4‑3.7× higher throughput, 29‑32% less GPU memory
Read on huggingface.coDetails
Hugging Face18 Jun 2026Fine-tuning
Beyond LoRA: Can you beat the most popular fine-tuning technique?
Problem Developers need to fine‑tune large models on their data without needing huge memory.
Solution They created a library that supports many parameter‑efficient fine‑tuning methods, letting users train only a small set of added weights and even work with quantized models.
Read on huggingface.coDetails
Etsy26 May 2026Fine-tuning
Shaping Product Understanding with Contrastive Reinforcement Learning
Problem Buyers often have vague tastes, making it hard to recommend the right handmade items.
Solution Etsy trained a model with contrastive reinforcement learning to better capture subtle product details, improving search and recommendation relevance.
Research behind this, on AI Papers DailyFactorized MLLM Embeddings for Composed Image Retrieval

Read on etsy.comDetails
Instacart4 May 2026Fine-tuning
Empowering Carrot Ads with Domain Adaptive Learning
Problem New partners had almost no ad interaction data, making it hard to predict which ads users would click.
Solution Instacart moved a model trained on its large marketplace to the partner’s site, then fine‑tuned it with the partner’s few clicks, giving better ad rankings.
Read on tech.instacart.comDetails
Hugging Face29 Apr 2026Fine-tuning
Granite 4.1 LLMs: How They’re Built
Problem Need small, high‑quality language models that can handle very long texts.
Solution Created 3B‑30B dense decoder models with a five‑phase pre‑training pipeline, curated data, supervised fine‑tuning using an LLM‑as‑judge, and reinforcement learning to boost math, code, and chat abilities.
8B model matches 32B predecessor
Read on huggingface.coDetails
Mercado Libre8 Jan 2026Fine-tuning
The silver shotgun: A bunch of silver bullets
Problem Income estimates varied across cities and neighborhoods, hurting credit decisions.
Solution They split data by location and user type, train separate models for each slice, and blend the outputs into one prediction.
Read on medium.comDetails