Machine Learning Librarian @huggingface 🤗 I like big datasets and small models.
Scotland
Joined September 2014
- Tweets5.2K
- Following1.5K
- Followers6.7K
- Likes9K
Pinned Tweet
Uploaded a dataset of 115,293 illustrated pages from the Encyclopaedia Britannica, 1st edition (1768) to 14th (1929) to the Hub
huggingface.co/datasets/bigl…
Daniel van Strien retweeted
Introducing Kimi K3.1 Pro: our most capable 3T-class open-weight model.
🔹 1M-token context, 2.8T total / 104B active params (MoE), Multimodal.
🔹 Full-project coding: games in Unreal, Unity & Godot.
🔹 Excels at parametric 3D modeling in Blender & CAD.
🔹 More reliable at finance work: Accurate modeling with hallucination down by 65%.
🔹 Matching the strongest proprietary models across coding, agentic and knowledge benchmarks.
Kimi K3.1 Pro and K3.1 Pro Highspeed (2–3× faster) are now live on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API.
Open weights by Nov 4, 2026.
🔗 API: platform.kimi.ai
🔗 Tech blog: kimi.ai/blog/kimi-k3-1-pro
Daniel van Strien retweeted
Everyone is talking about 'System One models' like Jev: given a state, question, and set of options, they return a calibrated decision at a fraction of the cost of an LLM call.
But do we need specialized architectures and recipes to train them?
We show you don’t:
🧵👇
After a month or so of open Jev-style decision models, here's what I found on the question of data.
Most of the findings come from one paper, "Releasing the Inner Jev" (@JonathanGeuter et al.).
You may need less training than you think. Show the options as letters, run one forward pass, and read the scores for "A", "B", "C"… An instruct model done this way already gets close to Jev: Gemma 4 31B scores 79.6 on OpenDecisions vs Jev's 80.1.
LoRA vs the head (31B ablations):
→ Train only the head, backbone frozen: 78.9 (slightly below zero-shot)
→ Train only LoRA, head frozen at the letter rows: 84.9
→ Full recipe, LoRA + head: 84.4
So the gain comes from adapting the representations. The letter readout is already a good head.
Also:
→ Short runs win (400–800 steps); longer runs mostly lose calibration
→ Smaller models gain less: no gain at 4B without extra tricks
→ A validation split from your own training data predicted the leaderboard badly
I also checked ~100 decision-related datasets on the Hub:
→ Most training data is classic classification/NLI datasets recast as choice/yes-no/score questions
→ Much of the rest is labelled by an LLM teacher
→ Eval sets were tiny (100–350 questions) until OpenDecisions (2,352)
→ Labels need re-checking: when Contested Fast Decisions relabelled the items 8 models disagreed on, 52% changed
I'm working on this a bit more to understand what data actually helps when you train one for your own task. So far I've reproduced the paper's zero-shot 4B result on @huggingface Jobs (64.7 vs 63.6, ~$0.20).
Paper: huggingface.co/papers/2610.1…
Datasets: huggingface.co/collections/d…
Daniel van Strien retweeted
Super excited to announce the @huggingface extension for DuckDB 2.0 !
Access (one of) the most optimized cloud storage in the world with hf://
Available day 1 on DuckDB 2.0
👉 hf.co/duckdb-extensions
Both @duckdb and @DataPolars got 2.0 releases this month. Which is faster for filtering big datasets on @huggingface Jobs?
Test: filter 286 GB of FineWeb-Edu by educational score and length.
→ 32 cores / 256 GB: Polars 352–447s, DuckDB 564–600s
→ 8 cores / 32 GB: DuckDB 1,904s (~2¢), Polars ran out of memory
Defaults, no tuning. There might be user/agent error! Scripts + job logs: huggingface.co/buckets/davan…
Daniel van Strien retweeted
BOOM! I trained my own judge on the Jev datasets. It beat Jev 1.14 🔥
Open 4B, 22 min on one NVIDIA H200 on Hugging Face Jobs, $1.87
4,035 held-out RAG + safety calls:
Jev 1.14: 82.2%
my 4B: 89.4%
Twist: on PubMed answers it never saw, Jev wins. Unless it learns from Jev 👀
New in @huggingface Jobs docs: a Train Models page.
A quick overview for you or your agent, covering Transformers, TRL, @UnslothAI, @axolotl_ai and Diffusers. Each example runs in minutes on a single A10G for around $0.10.
huggingface.co/docs/hub/jobs…
Daniel van Strien retweeted
I've been training OCR models for 2 years. The progress we've made in handwriting is astonishing - from barely readable to almost perfect transcription.
Daniel van Strien retweeted
Hugging Face Jobs usage is stonking 📈
All thanks to agents: you just need the hf CLI installed and you have access to an insane amount of hardware.
Asked mine to test hundreds of code snippets from Hub model pages: it launched 322 Jobs in 90 minutes, up to 25 at once, CPU to A100 (diffusers, transformers, sklearn, MLX, Keras...)
Total bill: ~$4 🤯
Daniel van Strien retweeted
1M trainings/month with TRL. 🤯
That’s what our telemetry is showing today.
Next stop: 1M/week. 👀
Every model you post-train, is one more step toward a less centralized AI ecosystem. Let’s go! ✊
Uploaded a dataset of 98,877 historical newspaper pages (1700s–1940s) to the Hub, each with its original OCR text, word boxes and confidence scores
huggingface.co/datasets/bigl…
Daniel van Strien retweeted
Announcing d1, our first decision model.
It's the first model to outperform Jev on @huggingface's Decision Index.
> wins on multilingual evals
> more robust against prompt injection
> handles longer inputs more effectively
> built for fast, structured decision-making in software environments
> Liquid API: console.liquid.ai
> Available on OpenRouter soon
Shipped datatrove 0.10.1 🛠️ Mostly a bug-fix release plus CI and testing work (Hypothesis property tests, faster CI).
github.com/huggingface/datat…
Daniel van Strien retweeted
Who wants more speech training data?
YODAS v3 is now available @huggingface
It’s 1.1M hours - the biggest audio dataset ever. And the first at this scale with stereo audio at 48kHz, with timestamped transcripts and translations. 100+ langs
hf.co/blog/espnet/yodasv3
CC-BY-3.0
Daniel van Strien retweeted
Hugging Face is where much of the machine learning community publishes and finds its datasets, while DuckDB is the in-process analytical database that queries files like CSV and Parquet directly, with no server or warehouse to install or run. 🤗 🦆
Did you know that, since #DuckDB v0.10.3 (~May 2024), you can point a SELECT at a dataset on the Hugging Face Hub, using the DuckDB `hf://` protocol, and query it, without downloading it first?
This blog post covers how that integration works and the use cases it fits: duckdb.org/2026/09/25/huggin…
The Hub can be a firehose of new releases.
I used @fastinoAI's GLiNER2.5-Decide to build a Space that tags every new model, dataset and Daily Paper on @huggingface into 48 topics, usually a minute or two after it changes.
Runs on one CPU, no LLM.
huggingface.co/spaces/davans…
Can now fine-tune an even better starter model using this recipe! xitter.cf/fastinoAI/status/21031…
Trained a Jev-style classifier on @huggingface Jobs for ~$1.50.
It's a 194M GLiNER2 model that suggests task tags for any Hub dataset from its column names and first row, and returns a label with a probability.
Zero-shot, GLiNER2's first suggestion matched an owner's tag 10% of the time. After 17 minutes of fine-tuning: 69%. The fine-tuned model runs on a free CPU in about a second.
Owners' tags are noisy, so some "wrong" answers are tags the owner left out.
The recipe is open: one hf jobs command trains the same kind of model on your own labels. The README example (book titles) runs in ~2 minutes for about $0.02.
Demo: huggingface.co/spaces/davans…
Recipe: huggingface.co/datasets/uv-s…