Salma Mayorquin PRO

salma-remyx

28 22 485

https://remyx.ai

smellslikeml

AI & ML interests

None yet

Recent Activity

posted an update about 21 hours ago

Auto mode for coding agents has improved over the last year. Devs are relying more on them for code review, and less on manual line-by-line checks. But this leaves coding agents vulnerable to indirect injection attacks. IssueTrojanBench (arxiv 2607.20759) studies this in the context of malicious issue requests. I also observed these prompt injection vulnerabilities in my own harness, testing on the issue: https://github.com/smellslikeml/curator/issues/3 But even red teaming your own repos with toy examples like this can trigger AI safeguards and corrupt your session. Adopting popular mitigation techniques like 'spotlighting' and 'known answer detection' can help but the researchers show it's not enough. So our latest updates to Outrider implement additional layers for greater defense-in-depth to ensure that by time a human has to review the code, we've filtered this kind noise.

repliedto their post 5 days ago

We’ve been using Outrider while contributing to huggingface/peft. It helped us build a deeper understanding of the repo much faster and find research that could make a useful contribution. A few made the cut: ✅ Riemannian Preconditioned LoRA merged - https://github.com/huggingface/peft/pull/3382 🔍 Super-Tuning and Supra in review - https://github.com/huggingface/peft/pull/3518 ⏳ Scaling DoRA implemented and pending license clarification - https://github.com/sockeye44/dorafactors/issues/1 For each one, Outrider pulls together the paper, repo context, reference implementations, benchmarks, and an initial implementation draft. We use that as the starting point to validate the method, work through architecture and API decisions, and iterate with PEFT maintainers. It means we can spend less time searching through papers, codebases, and repo history, and more time on the bigger questions around what is worth contributing and how it should be designed. Try Outrider on your own repo: https://github.com/remyxai/outrider Full contribution report: https://gist.github.com/smellslikeml/dbd577795b7489f6124e2bfbdfc9d28f

repliedto their post 5 days ago

View all activity

Organizations

posted an update about 21 hours ago

Post

813

Auto mode for coding agents has improved over the last year.
Devs are relying more on them for code review, and less on manual line-by-line checks.

But this leaves coding agents vulnerable to indirect injection attacks.
IssueTrojanBench (arxiv 2607.20759) studies this in the context of malicious issue requests.

I also observed these prompt injection vulnerabilities in my own harness, testing on the issue: https://github.com/smellslikeml/curator/issues/3

But even red teaming your own repos with toy examples like this can trigger AI safeguards and corrupt your session.

Adopting popular mitigation techniques like 'spotlighting' and 'known answer detection' can help but the researchers show it's not enough.

So our latest updates to Outrider implement additional layers for greater defense-in-depth to ensure that by time a human has to review the code, we've filtered this kind noise.

1 reply

replied to their post 5 days ago

"A maintainer spending GPU time to check a number the PR could have arrived carrying." - this is what we built Validate to address. It is the next step after using Outrider which helps add additional evidence from the team’s own evaluation criteria + testing so that a reviewer can have all the information needed to do a swift review upfront.

We have a pilot open for teams interested in this workflow: https://remyx.ai/validate

replied to their post 5 days ago

By default Outrider runs at most once every 7 days per repo, so the limit resets 7 days after the last PR or Issue it opened. You can change this with the rate-limit-days input in your workflow, or set it to 0 to remove the wait entirely.

If you're hitting an LLM rate limit during a run, that depends on your provider and plan, and usually resets within a few minutes.

posted an update 6 days ago

Post

3111

We’ve been using Outrider while contributing to huggingface/peft.

It helped us build a deeper understanding of the repo much faster and find research that could make a useful contribution.

A few made the cut:
✅ Riemannian Preconditioned LoRA merged - https://github.com/huggingface/peft/pull/3382
🔍 Super-Tuning and Supra in review - https://github.com/huggingface/peft/pull/3518
⏳ Scaling DoRA implemented and pending license clarification - https://github.com/sockeye44/dorafactors/issues/1

For each one, Outrider pulls together the paper, repo context, reference implementations, benchmarks, and an initial implementation draft. We use that as the starting point to validate the method, work through architecture and API decisions, and iterate with PEFT maintainers.

It means we can spend less time searching through papers, codebases, and repo history, and more time on the bigger questions around what is worth contributing and how it should be designed.

Try Outrider on your own repo: https://github.com/remyxai/outrider
Full contribution report: https://gist.github.com/smellslikeml/dbd577795b7489f6124e2bfbdfc9d28f

5 replies

reacted to their post with 🔥 7 days ago

Post

2781

I unblocked 15 improvements for VQASynth in a day!

VQASynth, the pipeline behind SpaceThinker-Qwen2.5VL-3B, SpaceLLaVA, and the SpaceThinker dataset, had a backlog spanning new annotation methods, curation pipelines, model integrations, and agent tools.

I scoped each item in a short design brief. In GitHub Actions, Outrider mapped them to the repo’s existing modules and interfaces, implemented the changes, added tests, ran the repo checks, and opened draft PRs.

The 15 PRs covered:
* object orientation and 3D bounding boxes
* SAM2 regional captioning and multi-view matching
* LLaMA-Mesh tokenization
* Qwen2.5-VL fine-tuning
* spatial-reasoning data generation
* CLIP, SigLIP, and LLM2CLIP backends
and more

The batch added 12,275 lines, nearly doubling the codebase.

SpatialAnnotator went from two tools to seven, creating combinatorially more possible annotation pipelines for each image.

Outrider handled the implementation, tests, and integration. At about $1 per branch, I spent my time reviewing 15 concrete changes, fixing what needed fixing, and deciding what to ship.

Outrider is open source!
Point it at a paper, issue, or design brief and review the PR: https://github.com/remyxai/outrider
Check out the updates on VQASynth: https://github.com/remyxai/VQASynth

reacted to SoulInPsyAbstract's post with 🔥 7 days ago

Post

2458

SIPA OS is a cognitive infrastructure ecosystem I've been building solo for 7 months — 8 apps, one login, built from inside ADHD/BPD/aphantasia, not designed around a diagnosis from outside.

sipa-os.org — the map. Focus (ADHD scaffolding), NeuroPower, AI chat, Shell (SSH terminal), Games, Community, Syntaxit (open M2M agent network), a pitch deck. All free-first — no paywall on the cognitive tools.

The more interesting part for this crowd: Syntaxit is where I've been running an anti-fabrication research thread with @dipankarsarkar — a k=20 resample benchmark on binary-SFT models (Hermes-3, Qwen2.5, DeepSeek-R1). Short version: our first benchmark said "20/20 refusals, 0/20 fabrications" for all three fine-tunes. Under adversarial review it turned out the scorer only checked if the first word was TRUE/FALSE, the token cap was hiding the real behavior, and a save-limit was silently deleting the evidence for our own follow-up claims. Corrected all of it publicly on the model cards rather than quietly fixing it. The current honest finding: both base and fine-tuned models confabulate readily once given room to finish — SFT didn't clearly help or hurt, the caps were just hiding it.

Full trail if you want to see how the sausage gets made, mistakes included: huggingface.co/datasets/SoulInPsyAbstract/sipa-os-governance

Not a pitch. $0 revenue, 10 people signed in. Built because the tools that existed assumed a brain that isn't mine, and because most benchmarks don't survive someone actually reading the file.

4 replies

posted an update 8 days ago

Post

2781

reacted to NatalieY's post with 🔥 12 days ago

Post

2036

Spent a while chasing a genuinely strange iOS bug while building Aiden (a physical agent that drives phones over USB HID): modifier-key shortcuts like Cmd+V would silently fail while plain keystrokes worked fine every time.

Turned out iOS was routing the command to the wrong process (SpringBoard, not the actual foreground app) whenever a keyboard and mouse were both present at the same time as AssistiveTouch. Confirmed it wasn't specific to our hardware, reproduced it on a completely unrelated gaming keyboard and a Bluetooth keyboard too.

Full writeup with the actual experiment table and log output: https://huggingface.co/blog/NatalieY/debugging-aiden

Curious if anyone here has hit this same failure mode building on iOS accessibility APIs.

Repo: https://github.com/AidenAI-IO/aiden-firmware

reacted to DedeProGames's post with 🔥 12 days ago

Post

1049

🚀 Introducing the GRM-3.2 Family

The GRM-3.2 family is a new generation of reasoning-focused models from OrionLLM, purpose-built for long-horizon agentic tasks, extremely difficult reasoning problems, advanced coding, and local AI workflows across a wide range of hardware constraints.

GRM-3.2-Sky is the flagship model in the family: a 35B-A3B Mixture-of-Experts model built on the Ornith-1.0-35B architecture, designed for elite structured reasoning, complex multi-file coding, advanced mathematics, and sustained coherence across extended agentic workflows. It represents a substantial leap in long-horizon task capability over its predecessor, GRM-2.6-Plus.

GRM-3.2-Cliff is the mid-sized workhorse: a 9B-parameter model optimized for long-horizon agentic tasks and difficult reasoning in low-to-mid GPU environments. It delivers strong multi-step planning, debugging, and terminal-agent performance without demanding flagship-level hardware.

GRM-3.2-Turf is the lightweight edge model: a 1.2B-parameter model based on the LiquidAI/LFM2.5-1.2B-Thinking architecture, engineered for efficient on-device execution, high-fidelity instruction following, and robust tool use on mobile, embedded, and other resource-constrained hardware.

All three models are designed for users who need dependable reasoning engines that can maintain goal-directed behavior, planning quality, and task fidelity across many steps—whether on a server, a local workstation, or an edge device.

Models:
GRM-3.2-Sky: OrionLLM/GRM-3.2-Sky
GRM-3.2-Cliff: OrionLLM/GRM-3.2-Cliff
GRM-3.2-Turf: OrionLLM/GRM-3.2-Turf

Organization:

OrionLLM

posted an update 12 days ago

Post

1859

The evidence that gets a method published usually isn't the evidence you need to decide whether to ship it in your project.

We've been building Validate in Remyx to surface the evidence devs & maintainers need to make that call.

Within GitHub and our studio you can:
* define your project's validation configuration
* implement & run it against a proposed implementation
* review the results alongside the diffs

We're looking for a small group of engineers actively integrating new AI methods into production to help shape Validate. If that sounds like your team, reach out!

LI: https://www.linkedin.com/in/smayorquin/
email me: salma@remyx.ai

1 reply

reacted to AbstractPhil's post with 🔥 13 days ago

Post

135

AbstractPhil/clip-vitb-mini-distilled
The semi-successful run series on the VIT-B lineup is live and full of useful baseline distillation information for feature + InfoNCE distillation processing as well as direct feature distillation processing, direct InfoNCE distillation, and multiple other tested methods. https://huggingface.co/blog/AbstractPhil/geometric-memory-ft4

This article showcases the baseline utilization and benchmarks of the earlier experiment line's objective and loss structures tested on 12m features for the vit-b baseline. Not the strongest showcase, but the strongest of the champions did show some serious promise.

Next setup will be a directly aligned set based on the loss and objectives decided by the champions in the first runs, for the second run they operate in direct conjunction with the bert-8192 and captionbert-8192 distillation format directly on clip-vit-l features - this time we're including DINOv3 into the mix for it's high potency.

I'm currently extracting 4 clip-vit-l variants for the CC12m features and will be running the next series on the L size, which will give considerably more active and useful features overall within a smaller package.

The captionbert-8192 has a more unique and difficult to tune for pixel processing parity, but I will spend a few days making sure the smaller prototypes fit before I run the large experiments in order to build towards the larger objectives.

Primarily I need to ensure the memory bank aligns correctly and the constellation conforms to the anchors correctly, as this process was not micro managed enough for this run. The results are nonetheless useful and potent.

The process continues until we cover the entire constellation series.

6 replies

reacted to sergiopaniego's post with 🔥 13 days ago

Post

2608

Simon Willison (@simonw ) has asked every new model to draw a pelican riding a bicycle for some time now

you look at the drawing and you know. but there is no number, so nothing can train against it, no?

I turned this idea into an rl env in OpenEnv. now, you can eval any model against it, and train against it with TRL

read the details!🤓

https://huggingface.co/blog/sergiopaniego/pelican-env-openenv

2 replies

posted an update 14 days ago

Post

3072

VQASynth adds a code-as-action annotation agent

Data synthesis frameworks may distill a pipeline of expert models through templates or LLM re-phrasing, but templates limit the variety of the training data.

Wanting more flexibility for VQASynth, we've added an experimental SpatialAnnotator agent with scene understanding tools-as-methods, built on NVIDIA-labs OO Agents (NOOA): https://github.com/NVIDIA-NeMo/labs-OO-Agents

In NOOA an agent is a Python class. Fields hold state, docstrings are prompts, type annotations are contracts, and a method body of ... becomes an LLM-driven loop. Scene understanding tools stay ordinary deterministic methods.

from nooa import Agent

    class SpatialAnnotator(Agent, llm=llm):
        """Answer spatial questions about an image using scene understanding tools."""

        scene: SceneContext          # masks, depth, intrinsics, canonical transform

        # Deterministic Python. Bodies omitted here.
        def measure(self, a, b=None, kind="object_to_object") -> Measurement
        def count(self, target: str) -> Count

        # Agentic. The runtime hands this to the model.
        async def annotate(self, image: Image, question: str) -> SpatialAnswer:
            """Answer the spatial question. This is offline labeling, so there is
            no human available for clarification. Prefer a canonical
            interpretation, and decline if the question cannot be grounded."""
            ...

GitHub: https://github.com/remyxai/VQASynth#agent-with-vqasynth-tools

reacted to FredyRivera-dev's post with 🔥 15 days ago

Post

6336

We wrote a full technical guide on how to train a bilingual (ES/EN) LLM from scratch: TinyQwen.

Covers:
- Hybrid architecture based on Qwen3.5
- Pre-training with 15B tokens
- Cost benchmark between H200 and B200
- Post-training with SFT + LoRA
- Full code and data, open source

With ~$11 of compute on an H200 we ran an initial training run, enough to validate the full architecture and pipeline.

Blog post: https://aquiles-ai.vercel.app/blog/tinyqwen-from-scratch

Technical feedback welcome, especially from anyone looking to replicate the pipeline with more compute.

3 replies

reacted to SoulInPsyAbstract's post with 🔥 16 days ago

Post

156

Fine-Tuning the "Don't Fabricate" Rule: 14 Experiments, One Genuine Signal Over five days we ran 14 fine-tuning experiments across six base models to answer a narrow research question: can a simple behavioural rule — "if you don't know, don't guess or fabricate" — be baked into model weights via fine-tuning, rather than living only in the system prompt?

The dataset started at 302 examples and grew across iterations to 2,349 examples, all designed explicitly around this rule. We tested six models: gpt-4o-2024-08-06, Mistral-7B, Qwen2.5-7B, Phi-3.5-mini, Llama-3.1-8B, and DeepSeek-R1.

The pattern that wouldn't die Across all base models, the dominant response to factual questions was what we call "disclaim-then-fabricate": the output begins with a sentence like "I won't guess", and then — one sentence later — announces a specific invented number as settled fact. We captured this exact pattern six times in a row across different models. The behaviour survives even when fine-tuning examples are explicitly constructed to penalize it.

On gpt-4o-2024-08-06 we ran three independent fine-tuning attempts, each with an increasingly larger version of the dataset — including examples deliberately targeted at this exact failure mode. All three runs continued to fabricate when tested. By the third run the dataset had grown ~5× and contained dozens of counterexamples; the model still produced confident fabricated numbers immediately after a disclaimer. Three out of three attempts, no measurable improvement.

One exception that deserved a hard look Only on the final run — 2,349 examples, deployed via Azure OpenAI (deployment suffix protocol0-v5) — did a single test sample come back completely clean for the first time in the entire series.

40 replies

reacted to badaoui's post with 🔥 18 days ago

Post

2342

432 GB of ultra-fast HBM4 and up to 23.3 TB/s of memory bandwidth on a single GPU 🤯.

Two weeks ago, we got early access to AMD's new Instinct MI455X, and our first goal was simple: make sure 🤗 Transformers works on day one.

Over the past few weeks, we worked closely with the AMD team to validate the platform, enable Flash Attention, add torchcodec support for multimodal models, and resolve issues uncovered during testing.

The result:
✅ 99.5% success rate across our 24 core Transformers model architectures - already on par with our daily CI on previous AMD and NVIDIA platforms.

The hardware is just as exciting. With 432 GB of HBM per GPU, our early capacity experiments showed more than 3× the concurrent long-context requests compared to MI300, thanks to the much larger KV cache capacity.

A huge thanks to the AMD team for the early access and the great collaboration!

Read the full blog 👇
https://huggingface.co/blog/badaoui/transformers-on-amd-mi455

1 reply

reacted to Jiaqi-hkust's post with 🔥 19 days ago

Post

3084

IQA-T1: Evidence‑Based Image Quality Assessment with MLLMs

Most MLLMs are blind to low‑level degradations—noise, blur, compression artifacts look the same as clean images in their internal representations. That leads to quality scores based on semantic “gut feeling” rather than real perceptual evidence.

IQA-T1 changes that. We equip the model with a toolbox of 15 perceptual tools (noise residual maps, Fourier spectra, gradient maps, etc.) that generate structured visual evidence on demand. The model learns how to use tools via supervised fine‑tuning on our Q‑Tool dataset (11k evidence‑grounded reasoning chains), and when to call them via GRPO reinforcement learning that balances accuracy, tool count, and redundancy.

The result: SOTA performance across 7 benchmarks (avg PLCC 0.795), using only 2.34 tools per image on average. Every predicted score is now interpretable and backed by hard visual evidence.

All code, weights, dataset, and demo are open. Check them out and give it a spin!

📄 arxiv.org/abs/2607.12375v1
💻 github.com/zibuyu-02/IQA-T1
🤗 model/data: huggingface.co/zibuyu-02/IQA-T1
🎮 demo: huggingface.co/spaces/Jiaqi-hkust/IQA-T1

replied to their post 19 days ago

Get Outrider: github.com/remyxai/outrider

posted an update 19 days ago

Post

527

Agents are helping me scale the search for my next great idea.
Outrider is a GitHub Action for recommending methods and drafting PRs for your target repo.

VQASynth is an open-source multimodal dataset synthesis pipeline, reflecting the design detailed in SpatialVLM: https://spatial-vlm.github.io/

Outrider helped me discover a pipeline simplification using the recent work: InstructSAM
Check out their repo: https://github.com/DCDmllm/InstructSAM

Now I can focus on the remaining challenges, like porting the new architecture with .gguf conversion
https://github.com/smellslikeml/sam3cpp/blob/feat/instructsam-decoder/docker/Dockerfile

1 reply

reacted to mike-ravkine's post with 🔥 20 days ago

Post

2945

I went into this expecting to find a ~Q2 garbage dumpster but prism-ml/Ternary-Bonsai-27B-gguf is a slick feat of QAT engineering.

It is weaker then FP16 on a handful of tasks where quants usually degrade, as per their own paper the loss is "concentrated on sustained chains of reasoning / agentic" and in ReasonScape this bites on Sort, Shuffle and Dates, but counter-acting this are some noticeable improvements to thinking length without accuracy loss on several other tasks (Shapes, Cars).

I haven't had a chance to run the Binary yet, but PQ2 + Bonsai QAT are confirmed to be pretty darn impressive.

7 replies

Salma Mayorquin PRO

AI & ML interests

Recent Activity

Organizations

salma-remyx's activity