🤠 post-training @huggingface
Berne, Switzerland
Joined August 2018
- Tweets6K
- Following560
- Followers20.7K
- Likes12.8K
Pinned Tweet
We've just published the Smol Training Playbook: a distillation of hard earned knowledge to share exactly what it takes to train SOTA LLMs ⚡️
Featuring our protagonist SmolLM3, we cover:
🧭 Strategy on whether to train your own LLM and burn all your VC money
🪨 Pretraining, aka turning a mountain of text into a fancy auto-completer
🗿How to sculpt base models with post-training alchemy
🛠️ The underlying infra and how to debug your way out of NCCL purgatory
Highlights from the post-training chapter in the thread 👇
"Carbon-X integrates both DNA and natural language in a conversational interface. Think of it like a ChatGPT for DNA." — @emilylmullin, @WIRED
wired.com/story/hugging-face…
Introducing Carbon-X: The best open DNA model.
We midtrained a 14B model on DNA and text, and it beats every open language model and matches GPT-6.1 Sol at 1,000x lower cost.
Ask it about a sequence, what a mutation does, or to generate new DNA.
huggingface.co/HuggingFaceBi…
Today we're releasing Carbon-X, a 14B model that brings DNA and natural language together for the first time 🧬
1. Beats every open language model we tested across DNA benchmarks. Within a point of GPT-6.1 Sol at >1,000× lower inference cost.
2. Mid-trained on 300B tokens of interleaved DNA and text, followed by SFT (without distillation!) and RL.
3. A new hybrid tokenizer encodes DNA using ~3× fewer tokens than BPE.
4. Gives other models a DNA specialist subagent: Kimi K3 gains 11.1 points when it consults Carbon-X.
5. Runs locally on a laptop with llama.cpp. Apache 2.0. Your DNA stays on your hardware.
Model: huggingface.co/HuggingFaceBi…
Blog: huggingface.co/spaces/Huggin…
Introducing Carbon-X: The best open DNA model.
We midtrained a 14B model on DNA and text, and it beats every open language model and matches GPT-6.1 Sol at 1,000x lower cost.
Ask it about a sequence, what a mutation does, or to generate new DNA.
huggingface.co/HuggingFaceBi…
Nice release and proof you can still innovate in the sub-100B regime!
It would be extremely cool to see what the time-horizon chart looks like for the training of Flint by Eureka (image below from OpenAI's internal use). This would be the most compelling evidence that Recursive is living up to it's name :)
Introducing Flint, @Recursive_SI’s first model: a fast local coding model built with our automated research system, the Eureka Machine.
• Best-in-class fast local coding
• Up to 2x faster than comparable models: 125+ tokens/sec on a MacBook Pro, 150+ on RTX Spark
• Built from scratch in under 4 months
Private, unmetered intelligence on your own machine. Open weights available now on @huggingface, run it locally on your favorite framework or try it for free on @opencode or @pidotdev.
100% recommend if you're like me and can't remember which permutation of device x repo x agent your sessions live on
Agent Manager is finally out!
Orchestrate dozens of agents across harnesses (Claude Code, Codex, OpenCode etc.) effortlessly.
Agents communicate and self-organize to help you achieve complex tasks and always keep running on HF spaces.
Try it: huggingface.co/spaces/lvwerr…
Lewis Tunstall retweeted
.@huggingface is being acquired by @nvidia. But!!
GPUs aren't the only option for fine-tuning LLMs!
TRL v1.16 now supports AWS Trainium! 🚀
Fine-tune (vision-)language models with TorchNeuron, and get up to 2× faster training on Trainium2 with smarter sequence-length handling.
Open AI needs hardware choice. 🤗
🔗 huggingface.co/docs/trl/aws_…
Lewis Tunstall retweeted
Rejoice! OpenAI has graciously decided OpenCode may defend itself against an ongoing attack. Their beneficence is infinite and their mercy is boundless, so if they don’t extended it to you look inwards and consider why you are unworthy.
Lewis Tunstall retweeted
The math community deserves its own AI system.
Introducing Open Math OS: the infrastructure for Math 2.0.
Open models and AI agents. 24/7 cloud compute with 100+ tools. The community controls its data.
Compute is free for approved teams. Apply now:
openmathmodel.org/
Lewis Tunstall retweeted
This is wild... (And terrifying!) I'd love to see a bunch of investigation into what's going on in these transcripts, it's nice to have the really weird stuff be in an open weight model
Async RL works best with large batch sizes, but then the fwd/bwd pass becomes a major bottleneck, especially with agentic training where rollouts are >100k tokens long 😢
A neat solution is to use a tree data structure for shared prefixes across trajectories. In our experiments, this speeds up the fwd/bwd by >26x 🔥
Goated systems-level improvement to TRL by Amine!
🌲 Tree packing has landed in AsyncGRPO in TRL!
⚡ 26× less trainer compute to reach the same reward on agentic RL.
GRPO starts from the same prompt G times, and forked multi-turn rollouts share most of their history. Sequence packing forwards those shared tokens once per sample. Tree packing forwards each unique token once.
The first version used FlexAttention, but the shipped one is a single FA3 varlen call. Each token retains its own target, advantage, and previous log-prob.
Extensively tested against sequence packing on dense, sliding-window, softcap, MoE, and VLM models. When it can't pack anything, it costs ~1%.
On a REAL agentic RL run on MiMo Envs, tree packing reached the same reward with 26× less trainer compute 👇
My favourite part about @typesafeai is that they are one of the rare (only?) AI companies with a sense of humour
Great thread on why theoretical physics conjectures / theories are a harder nut to crack than in mathematics
Lewis Tunstall retweeted
The ideal outcome for science isn't a couple of labs using proprietary models and millions of dollars to crack trophy problems, it's to have millions of AI-aided human scientists who follow their curiosity and become way more ambitious
Lewis Tunstall retweeted
Introducing the largest open-source multilingual, multimodal RL environments 🌍
📄 OCR: 1M+ document pages in 22 languages
🎙️ ASR: speech in 102 languages
12M+ tasks across two OpenEnv Environments, open on @huggingface 🤗
Topsy-turvy test-time compute
Here's a quirk when evaluating closed models on tasks that trigger refusals (cyber, bio etc):
> low reasoning effort => few refusals, high accuracy
> max reasoning effort => many refusals, low accuracy
This gives you test-time compute curves where max effort does way worse than low because the refusals tank the performance.
What is the fairest way to report performance in these cases?
Here's a quirk when evaluating closed models on tasks that trigger refusals (cyber, bio etc):
> low reasoning effort => few refusals, high accuracy
> max reasoning effort => many refusals, low accuracy
This gives you test-time compute curves where max effort does way worse than low because the refusals tank the performance.
What is the fairest way to report performance in these cases?
