Posts

A hand selecting a card from an index with labelled drawers.

Building data curation classifiers with agents, SetFit and Jobs

agents
datasets
hf-jobs
huggingface

Building a small document classifier with an agent, then using Hugging Face Jobs to label 1% of the English FinePDFs-Edu dataset.

Sep 10, 2026
Daniel van Strien
 

Distilling Qwen3.8 with datatrove on Hugging Face Jobs

35,000 card summaries for $16, no local GPU. Part 1: the data.
synthetic-data
hf-jobs
datasets
huggingface

datatrove’s new Jobs backend plus a days-old 27B teacher: regenerating a 35,837-summary training dataset in one afternoon for $15.58, with a calibration-first workflow and length-controllable outputs.

Aug 17, 2026
Daniel van Strien

A model release became a training recipe the same day. I mostly watched.

LiquidAI shipped new encoder models this morning. By the evening there was a fine-tuning recipe, a tutorial dataset, and a working classifier — built by an agent on HF Jobs for under $5 in GPU time.
agents
fine-tuning
hf-jobs
huggingface

I pasted two links into Claude Code and asked for a training recipe with early smoke tests on HF Jobs. What came back — and why agent-legible infrastructure is what made it work.

Jul 28, 2026
Daniel van Strien

Benchmarking OCR fairly is harder than it looks

Ten OCR models on olmOCR-bench’s hardest scans — and why a model that reads the page more faithfully can score lower.
ocr
glam
evaluation
hf-jobs

I ran ten current OCR models on olmOCR-bench’s ‘old scans’ subset. The ranking flips depending on one thing: whether you want a model that keeps what’s on the page, or one that cleans it up. A note on what the headline score actually rewards — and how agents + Jobs made it cheap to run.

Jul 7, 2026
Daniel van Strien

Is fine-tuning small models worth it? Structured extraction from index cards

Fine-tuning one open 4B model to read several archival card collections, each with its own schema — and still follow schemas it has never seen.
ocr
glam
structured-extraction
computer-vision
fine-tuning
hf-jobs

Can a small, open model, fine-tuned cheaply, match a bigger one on a real extraction task and stay general? A test on archival index cards across several collections, for about $45.

Jun 24, 2026
Daniel van Strien

Filtering 1 TB of FineWeb-Edu for $2.40 using Buckets and Jobs

data-engineering
polars
hf-jobs
huggingface

A quick look at using Hugging Face Jobs + Storage Buckets to filter a big dataset — DuckDB does the heavy scan, Polars picks up the result, a bucket is the handoff. ~1 TB filtered for about $2.40.

Jun 2, 2026

How to turn catalogue card images into structured JSON with a 4B open model

A 4B open model turns catalogue-card images into schema-shaped JSON — the kind of structured data that can be ingested back into a library system.
ocr
glam
structured-extraction
datasets
huggingface

Re-OCR made digitised catalogue cards searchable as text. But libraries need structured records they can ingest into their catalogues. NuExtract3 (4B, Apache-2.0) extracts schema-shaped JSON from card images in one command — here’s how it does, honestly.

May 21, 2026
Daniel van Strien

All four agents tied at F1 ~0.95. The number was a mirage.

Three harnesses, two driver models, four runs. The traces show why an outcome metric on this dataset doesn’t mean what the agents reported it means — and a 22-point F1 gap when you actually test for generalisation.
agents
glam
datasets
huggingface

Same one-line prompt, three harnesses, two driver models. The headline F1 tied within 1.5pp. A source-stratified eval the agents didn’t run shows the headline number was hiding a 22 F1-point generalisation gap.

May 6, 2026
Daniel van Strien

Two agents, one prompt

Pi + Kimi K2.6 and Claude Code both fine-tuned a Jim Crow law classifier from a single sentence. The interesting gap wasn’t F1.
agents
glam
datasets
huggingface

I gave two coding agents the same one-line prompt and watched what each decided to do. Both shipped working classifiers via hf jobs. The headline metrics were within 1.5pp — the interesting differences were everywhere else.

May 1, 2026
Daniel van Strien
 

AI providers have millions of agent sessions. The first 1,589 are public.

A first pass at 8,949 labelled messages from the only open coding-agent corpus we’ve got — and why sentiment is probably the wrong lens.
agents
datasets
huggingface

Anthropic, OpenAI, Cursor, and Windsurf sit on millions of private agent traces. The first public corpus has 33 datasets, 1,589 sessions, 8,949 labelled messages. Here’s what we found in it — and why the interesting data isn’t in the words.

Apr 20, 2026
Daniel van Strien, Claude Opus 4.7
 

Re-OCR Your Digitised Collections for ~$0.002/Page

ocr
glam
huggingface

A guide to re-processing digitised collections with open-source VLM-based OCR models.

Feb 19, 2026
Daniel van Strien

Train on Massive Datasets Without Downloading with Hugging Face Streaming and Unsloth

huggingface
datasets
fineweb
fine-tuning

Stream datasets directly from Hugging Face Hub without downloading. Combine streaming with Unsloth for disk-free LLM training on Colab, Kaggle, or HF Jobs.

Jan 7, 2026
Daniel van Strien

Fine-tuning Vision-Language Models for Art History: Iconclass Classification with TRL and HF Jobs

huggingface
hf-jobs
computer-vision
art-history
fine-tuning

Learn how to fine-tune open-source VLMs like Qwen2.5-VL for specialized art history tasks using Iconclass metadata. This tutorial shows how to use TRL’s new VLM support with Hugging Face Jobs for cloud-based training - no local GPU required!

Sep 4, 2025
Daniel van Strien

Efficient batch inference for LLMs with vLLM + UV Scripts on HF Jobs

huggingface
vllm
hf-jobs

Generate responses for thousands of dataset prompts using Qwen/Qwen3-30B-A3B-Instruct-2507 across 4 GPUs with automatic prompt filtering and tensor parallelism

Jul 30, 2025
Daniel van Strien

Who Benefits? Rethinking Library Data in the Age of AI

libraries
datasets
glam

From data providers to AI partners: How libraries can collaborate with developers to build tools that improve collections, enhance discovery, and serve communities while contributing to responsible AI development.

Jun 9, 2025
Daniel van Strien

Efficient Inference for ModernBERT Classifiers Using vLLM

huggingface
datasets
vllm

Modern Inference for modern models. Using vLLM to scale ModernBERT based classifier models to clean and curate datasets

Apr 24, 2025
Daniel van Strien

Using QwQ to generate a reasoning dataset for structured data extraction

huggingface
datasets
synthetic-data
structured-extraction

Learn how to use QwQ-32B to generate synthetic reasoning datasets for training smaller models on structured data extraction tasks

Mar 11, 2025
Daniel van Strien

Log GRPO Completions to 🤗 Datasets

huggingface
datasets
fine-tuning

Log GRPO training completions from trl to a 🤗 Dataset repo for easy analysis

Feb 20, 2025
Daniel van Strien

Distiling DeepSeek reasoning to ModernBERT classifiers

huggingface
datasets
synthetic-data
fine-tuning

How can we use the reasoning ability of DeepSeek to generate synthetic labels for fine tuning a ModernBERT model?

Jan 29, 2025
Daniel van Strien

Hygge Data - Cozy Content Filtering for a finer Scandinavian FineWeb

polars
huggingface
fineweb
datasets

Training lightweight, disposable web scale data curation models for Scandinavian language texts using the FineWeb-c dataset

Jan 9, 2025
Daniel van Strien

How Fine is FineWeb2?

polars
huggingface
fineweb

Analysis of the FineWeb-C dataset

Jan 7, 2025
Daniel van Strien

Filtering FineWeb2 using Polars

polars
huggingface

Using Polars to filter the FineWeb2 dataset and other large Hugging Face datasets

Dec 30, 2024
Daniel van Strien

Running a Local Vision Language Model with LM Studio to sort out my screenshot mess

computer-vision
huggingface
local-ai

How I used a VLM to sort out my screenshot mess

Oct 23, 2024
Daniel van Strien

Using ColPali with Qdrant to index and search a UFO document dataset

computer-vision
huggingface

Using Qdrant multivectors to index and search a UFO document dataset

Oct 2, 2024
Daniel van Strien

Generating a dataset of queries for training and fine-tuning ColPali models on a UFO dataset

computer-vision
datasets
synthetic-data

Learn how to generate custom ColPali dataset using an open VLM for multimodal retrieval model training and fine-tuning.

Sep 23, 2024
Daniel van Strien
 

Synthetic dataset generation techniques: generating custom sentence similarity data

datasets
synthetic-data

This post is part of a series on synthetic data generation techniques. You may also want to check out Awesome Synthetic (text) datasets, where I will be collecting these posts.

May 23, 2024
 

Synthetic dataset generation techniques: Self-Instruct

datasets
synthetic-data

This post is part of a series on synthetic data generation techniques. You may also want to check out Awesome Synthetic (text) datasets, where I will be collecting these posts.

May 15, 2024

Tracing Text Generation Inference calls

huggingface
evaluation

How to trace text generation inference calls with Langfuse.

Apr 5, 2024
Daniel van Strien
 

Extracting Insights from Model Cards Using Open Large Language Models

huggingface

What do people talk about in their model cards?

Nov 27, 2023
 

How to load a Hugging Face dataset into Qdrant?

datasets
huggingface
data-engineering

Loading a Hugging Face dataset into Qdrant is easy. This post shows how to do it.

Nov 8, 2023
Daniel van Strien

How to do groupby for Hugging Face datasets

datasets
huggingface

How can you groupby in Hugging Face datasets?

Sep 18, 2023
Daniel van Strien

Exploring language metadata for datasets on the Hugging Face Hub

datasets
huggingface

Using the huggingface_hub library to asses metadata on the hub

Jun 7, 2023
Daniel van Strien

Dynamically updating a Hugging Face hub organization README

huggingface

Using the huggingface_hub library and Jinja to update a README dynamically

Mar 7, 2023
Daniel van Strien

Using Hugging Face AutoTrain to train an image classifier without writing any code.

huggingface
fine-tuning

How can we train useful machine learning models without writing code?

Feb 22, 2023

A (very brief) intro to exploring metadata on the Hugging Face Hub

huggingface
datasets

How we can use the huggingface_hub library to explore metadata on the Hugging Face Hub.

Jan 16, 2023
Daniel van Strien

Label Studio x Hugging Face datasets hub

huggingface
datasets

Using label studio and the Hugging Face datasets hub to iteratively annotate a dataset

Sep 7, 2022
Daniel van Strien

Training an object detection model using Hugging Face

computer-vision
fine-tuning
huggingface

Training a Detr object detection model using Hugging Face transformers and datasets

Aug 16, 2022
Daniel van Strien

Searching for machine learning models using semantic search

huggingface

Finding models on the Hugging Face hub using semantic search

Jul 26, 2022
Daniel van Strien

Combining Hugging Face datasets with dask

datasets
huggingface
data-engineering

Using 🤗 datasets in combination with dask

Jun 20, 2022
Daniel van Strien

Using 🤗 datasets for image search

computer-vision
datasets
huggingface

Using the 🤗 datasets to make an image search engine for British Library Book Illustrations

Jan 13, 2022
Daniel van Strien
 

Using the 🤗 Hub for model storage

huggingface

How I’m planning to use the huggingface hub for storing flyswot models

Dec 30, 2021
Daniel van Strien

flyswot

glam
computer-vision
fine-tuning

Attempting to deploy machine learning in an existing workflow

Dec 22, 2021

Image labeling vs classification models

computer-vision
datasets

Comparing the loss functions of label and classification models

Oct 12, 2020
Daniel van Strien

Hyperparameter Optimization for Transfer Learning

fine-tuning
computer-vision

Optimising Hyperparameters using optuna and fastai2

Jul 1, 2020
Daniel van Strien
 

Multi-model metadata generation

datasets
computer-vision

Experiment in combining text and tabular models to generate web archive metadata

May 3, 2020
Daniel van Strien
No matching items