@axolotl_ai OSS maintainer. Axolotl AI founder. AI/ML tinkerer. Building tools for everyone.
Annapolis, MD
Joined March 2009
- Tweets3K
- Following1.9K
- Followers11K
- Likes3.5K
Day-0 support LoRA fine-tuning Le Chonk (Mistral Large 4) is in Axolotl. Our recommended recipe is experts in NVFP4 using merge-aware LoRA. Fits on one 8xB200 (8xB300 is ideal).
Today, we release the weights of a new and improved Le Chonk, advancing the open frontier for agentic coding and cybersecurity.
💻 Built to build: Le Chonk ranks #1 on @cognition’s FrontierCode among open models, and sets a new benchmark for Pareto efficiency on @artificialanalysis.
🛡 Built to defend: State-of-the-art in cyberdefense, put through the paces by the world's leading cyber vendors, including @CrowdStrike, @cloudflare, @rubrik, @zscaler, and others.
Built for sovereignty: Make it yours. Download the weights on @huggingface , customize it with Mistral Forge, and deploy it wherever you need it.
Read more from our team and customers 👇:
NVFP4 weights: huggingface.co/mistralai/Mis…
Docs: docs.axolotl.ai/docs/models/…
If you're using LoRA for RL, it's most likely best to deploy your adapter unmerged. Depending on your learning rate, how many effective train steps, etc, you might be losing much of your training updates to rounding to BF16 when you merge the adapter to the base model.
In Axolotl, we’ve addressed this with merge-aware LoRA training: we merge the LoRA into the base weights and cast to BF16 inside the forward pass, so the model trains on exactly the weights vLLM serves. Gradients pass straight through the rounding to the LoRA weights, so they keep learning even when a single step is too small to change a BF16 weight. While there are still gaps in this method (support for various distributed parallelisms, ZeRO-3, etc), we hope we can help ensure that you aren't left wondering why what you deployed is underperforming.
If you're shipping the LoRA adapter to your rollout engine during training, you don't have the rollout-training mismatch since there is no rounding off. But if you plan to merge for deployment, the same rounding issue still applies. At the end of the day, the small differences may not matter, but something to be aware of.
You can now use AsyncGRPO with tree packing on your favorite hybrid models such as @Alibaba_Qwen 3.5 (GDN), @liquidai LFM 2.5, and @NVIDIAAI Nemotron 3.5 (Mamba2). Built as a standalone component that hooks into @huggingface Transformers and leverages Hub kernels.
🌲 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 👇
Works with vanilla TRL out of the box. @axolotl_ai integration coming soon. Check out our example to get started at github.com/axolotl-ai-cloud/…
Any good & free PR review bots out there for OSS projects? We are using @coderabbitai but they decided to start adding prompt injections so your agent will install their tool locally and use that to check. In the meantime we've had to include prompt injection mitigation in our CONTRIBUTIONS.md.
Wing Lian (caseus) retweeted
Today, I’m open-sourcing Drost-v1 as Drost Community Edition.
Drost Community is a genuinely powerful offensive cybersecurity tool: It comes with one Docker container, 159 MCP tools, Linux and MacOS support, and direct integration with MCP-compatible AI agents.
In short, you can use your existing coding agents to carry out offensive cybersecurity engagements.
It runs entirely locally, on your hardware, with your models. Drost receives no engagement telemetry.
This is where Drost began.
drost.ai/community
Enjoy! 🚀💥
We can show via experiments on synthetic tasks that there is may be an improvement on some tasks from balancing label counts. While this does show that packing may in fact be worse for this particular synthetic set of tasks, we do see a trend that optimizing for label counts does seem to help.
Axolotl had an issue a while back where we were doing BFD over the entire dataset and saw loss divergence near the end of an epoch where the short tail of multiple documents were concentrated. More modern BFD implementations pack over shorter windows to mitigate this, but I’m not sure structured packing like this is optimal.
But can we do better? For most SFT workloads, we’re coordinating two things, packing to minimize padding tokens, minimizing loss via supervised labels. The thing we’re not doing today is also co-optimizing the distribution of supervised labels. By co-optimizing for label counts we can sharpen the distribution of labels to reduce gradient noise across steps.
Using BFD, you end up with longer samples getting paired up with smaller samples. Depending on the correlation of tasks to sample lengths, this can lead to some pairing off of tasks within a batch (at smaller effective batch sizes).
Hot take for today; BFD packing is suboptimal for actual post-training. Your effective batch sizes are too small and your trading off structured packing for accuracy.
Just found an additional 8% speedup in axolotl just by adapting our existing fused rope+rms norm attention kernels for GLM4.
If you're curious why they only reported against TRL for their single GPU SFT on GLM Flash, it's because at that regime, they're still slower than Axolotl by about 5% (can't find any B300s, so using B200s). But hey, reporting benchmarks these days is just about cherry-picking data, right?
Introducing Halo, the best framework for post-training of open-source models.
Halo delivers up to 2.8x the throughput of stock TRL with less peak memory, while models stay in their native HuggingFace format.
Star us on GitHub: github.com/whitecircle/halo
Wing Lian (caseus) retweeted
After many months of intense work the @Snowflake AI Research team is happy to present to you the new open source project: Arctic RL
snowflake.com/en/blog/engine…
- Arctic RL integrates with VeRL and SkyRL today; enable ZoRRo with one config flag, no code changes required
- ZoRRo delivers up to 6x actor-update acceleration and a 3.5x end-to-end training speedup, reducing Arctic-Text2SQL-R2 training from ~5 days to ~36 hours on 32 H200 GPUs
- Arctic-Text2SQL-R2 achieved higher accuracy scores (48.7) than Gemini 3.1 Pro (47.9) and Claude 4.7 (47.3) on Snowflake's evaluated enterprise SQL benchmark under the tested conditions
- Two open source recipes ship with this release: a text-to-SQL recipe that improved BIRD dev accuracy from 59.92% to 70.35%, and a multi-hop QA recipe that improved average accuracy from 69.6% to 72.3%
I honestly have no words for calling it “megakernel” when in fact it’s just a bag of kernels.
Megakernel for Laguna XS + GRPO training on a single RTX 4090:
~180 tok/s @ batch 1
~499 tok/s @ batch 8
In-place, memory-efficient forward & backward wiith custom kernels.
@poolsideai @eric_alcaide
github.com/KONAKONA666/fastg…
