Trustpilot
Trustpilot fine-tunes small Gemma models to label millions of reviews a day
Trustpilot fine-tuned small Gemma models, Google's open-weight family, to label millions of customer reviews a day for topic, named entities and sentiment. Google Cloud and Trustpilot report accuracy a couple of percentage points below the larger models that labelled the training data, at a fraction of the cost.
Key figures
Copied exactly from the sources. Copy any figure with its citation.
Millions a day, in near real time
Reviews processed, as described by the authors1
A couple of percentage points lower
Accuracy of fine-tuned models against the consensus of the larger teacher models1
Considerably better (no figure given)
Accuracy against the previous solution1
What it took
The problem
Trustpilot's business depends on turning customer reviews into usable data: which topic a review covers, which companies or products it names, how positive it is and what the writer wants. It receives millions of reviews and needs the results quickly. Sending each one to a leading commercial model is costly at that volume, and the authors say a closed model also ties Trustpilot to a vendor's pricing and update schedule.
What they built
Trustpilot built a set of small, specialised models on Gemma 2 9B, an open-weight model from Google. Open-weight means the model's files are published, so a company can adapt it and run it on its own cloud account. Fine-tuning means further training a model on a company's own labelled examples. The models handle topic classification, named entity recognition (finding names of companies and products in text) and sentiment extraction.
How they did it
The team first had larger Gemini models from Google label a representative sample of Trustpilot reviews. It kept the labels the teacher models agreed on and used them as training data. These "teacher" labels trained the small models.
In production, a streaming pipeline on Google Cloud sends each review to two services. One handles preparation and prompts. The other serves Gemma through vLLM, an open-source serving tool, on A100 GPUs. The authors tuned vLLM settings, built a reusable load-testing method and set autoscaling by request count. The joint project with Google ran at the end of 2025.
Results
The authors report that the fine-tuned models clearly beat Trustpilot's previous solution and came within a couple of percentage points of the consensus of the teacher models. They say the pipeline processes millions of reviews a day in near real time, with performance like a Gemini model for a fraction of the cost. The post gives no figures for the cost saving.
What went wrong
Private networking between the two endpoints was not natively supported, so the fully isolated design was not possible. Deployments were slow and opaque, and A100 chips in the EU region were hard to secure.
Lessons for your project
Use a large model as a teacher, not as the production engine. Trustpilot had larger models label a sample of reviews, then trained small models on those labels.
Choose fixed infrastructure cost over per-token pricing when volume is high and steady. Trustpilot cites this predictability as a reason for open-weight models.
Keep the language model and the business logic in separate services so each can scale on its own.
Check GPU supply early. Securing A100 chips in the EU region was hard, so reservations had to be shared across development, training and production.
Caveats
- The post is co-written by Trustpilot and Google staff and gives no absolute figures for cost, latency or accuracy, only relative statements.
- The post does not say when the pipeline went live. The deployed date shown is the month the post was published; the joint project ran at the end of 2025.
- The post does not state the region. Europe is based on Trustpilot being a Danish company and the post's use of an EU computing region.
- The authors list unresolved issues: private networking between endpoints was not possible and endpoint deployments were slow and hard to observe.
Sources
- How Trustpilot built a real-time architecture for data enrichment using Gemma
Google Cloud Blog · 2026-06-02 · Vendor case study · cloud.google.com
Summarized in our own words from the sources above. Spotted an error? See corrections.