53 8 246

sometimesanotion

https://ko-fi.com/sometimesanotion

AI & ML interests

Agentic LLM services, model merging, finetunes, distillation

Recent Activity

liked a model about 2 hours ago

keithnull/Qwen3.6-35B-A3B-REAM-192-heretic-APEX-GGUF

liked a model 3 days ago

arcee-ai/Trinity-Mini

liked a model 5 days ago

google/diffusiongemma-26B-A4B-it

View all activity

Organizations

reacted to danielhanchen's post with 🔥 10 days ago

Post

8938

Gemma 4 12B can now run locally on just 8GB RAM via Dynamic GGUFs.

Google's new model, Gemma 4 12B Unified supports image, audio and 256K context.
You can run and train the model via Unsloth Studio.

GGUF: unsloth/gemma-4-12b-it-GGUF
Guide: https://unsloth.ai/docs/models/gemma-4

5 replies

reacted to sequelbox's post with 🚀 about 2 months ago

Post

1926

NEW RELEASE: Esper 3.1 for Qwen 3.6!

- Your dedicated DevOps expert: Esper 3.1 maximizes DevOps and architecture helpfulness, powered by high-difficulty DevOps and architecture data generated with DeepSeek-V3.1-Terminus!
- Improved coding performance: challenging code-reasoning datasets stretch DeepSeek-V3.1-Terminus and DeepSeek-V3.2 to the limits, allowing Esper 3.1 to tackle harder coding tasks!
- AI to build AI: our high-difficulty AI expertise data boosts Esper 3.1's MLOps, AI architecture, AI research, and general reasoning skills.

Get it now: ValiantLabs/Qwen3.6-35B-A3B-Esper3.1

We're working on more finetunes for the newest Qwen and Gemma models, and we've also started working on the agentic-first datasets for Esper 4 :) we're going to make open source better and better for your work!

Please note that real life financial and family concerns have popped up and have imposed unfortunate limitations on our ability to devote time to our open-source work :( If you would like to see Esper 4 and our other releases speed up instead of slowing down, this is the best way you can help us: sequelbox/SupportOpenSource

No matter what, we'll keep fighting and we won't give up!

with love,
allegra

1 reply

reacted to Nymbo's post with 👍 3 months ago

Post

7403

We should really have a release date range slider on the /models page. Tired of "trending/most downloaded" being the best way to sort and still seeing models from 2023 on the first page just because they're embedded in enterprise pipelines and get downloaded repeatedly. "Recently Created/Recently Updated" don't solve the discovery problem considering the amount of noise to sift through.

Slight caveat: Trending actually does have some recency bias, but it's not strong/precise enough.

3 replies

replied to DavidAU's post 4 months ago

Neat! I'd love to see one of these fine-tuned on a tool-calling dataset.

reacted to DavidAU's post with 👍 4 months ago

Post

6968

The "ERNIE" 21B MOE Distill High Reasoning Fine Tune Invasion:

3 Ernie 21B-A3B MOE Models (64 experts) fine tuned with Unsloth using Gemini Pro 3, Claude 4.5 Opus, and GLM 4.7 Flash high reasoning datasets.

All benched, all exceeding org model specs too.

https://huggingface.co/DavidAU/models?search=ernie

Enjoy the freedom and added power.

1 reply

reacted to sequelbox's post with ❤️👀 5 months ago

Post

2695

NEW RELEASE: it's here! Meet the newest member of the Valiant crew: Guardpoint, our new medical reasoning model!
- Trained on medical knowledge, management, diagnosis, and tasks from DeepSeek-V3.2-Speciale!
- Structured medical reasoning responses are efficient and informative, cutting token costs for faster inference!
- Wide-ranging knowledge base: trained on a wide variety of medical disciplines, patient types, and query structures!
- High quality medical responses emphasize performance, brevity, specificity, statistical rationality, and openness.

Get it now:
Guardpoint for Qwen 3 32B: ValiantLabs/Qwen3-32B-Guardpoint
Guardpoint for Qwen 3 14B: ValiantLabs/Qwen3-14B-Guardpoint
Powered by our new structured medical reasoning dataset: sequelbox/Superpotion-DeepSeek-V3.2-Speciale

We've been working hard on Guardpoint; we're really excited to share it with everyone!

We'll be bringing Guardpoint to more models soon, along with further releases for the Shining Valiant and Esper series!

Get our experimental models: https://huggingface.co/collections/sequelbox/experimental-reasoning-models
Get our reasoning datasets: https://huggingface.co/collections/sequelbox/reasoning-datasets

Help support our releases, donations used for our experimental models and datasets: sequelbox/SupportOpenSource

2026 is going to be an amazing year for open source AI! It's time for the AI revolution you need; from the bottom up, built together by all of us.

for love, friendship, and better days,
allegra

1 reply

reacted to csabakecskemeti's post with 🔥 6 months ago

Post

3381

Just sharing a result of a homelab infrastructure experiment:

I've managed to setup a distributed inference infra at home using a DGX Spark (128GB unified gddr6) and a linux workstation with an RTX 6000 Pro (96GB gddr7) connected via 100Gbps RoCEv2. The model I've used (https://lnkd.in/gx6J7YuB) is about 140GB so could not fit either of the GPU. Full setup and tutorial soon on devquasar.com

Screen recording:
https://lnkd.in/gKM9H5GJ

3 replies

reacted to sequelbox's post with 🔥 6 months ago

Post

3045

Two new releases today!

Firstly, our new Raiden-Mini dataset, powered by DeepSeek's newest deepseek-ai/DeepSeek-V3.2-Speciale model!
- A V3.2-Speciale reasoning showcase: the Raiden prompts test the model's creative, analytic, and general reasoning skills!
- HEAD TO HEAD: a comparison subset pits V3.2-Speciale against V3.2 with the same prompts, providing a direct look at each model's advantages!

Get the new Raiden-Mini dataset: sequelbox/Raiden-Mini-DeepSeek-V3.2-Speciale

On the model side, we've also brought Shining Valiant 3 to Ministral 3!
- Science-reasoning: sequelbox/Celestia3-DeepSeek-R1-0528 for physics, biology, chemistry, compsci, astronomy, Earth science, and information theory.
- AI to build AI: the sequelbox/Mitakihara-DeepSeek-R1-0528 dataset for high-quality reasoning performance on AI, MLOps, math and CUDA, complex adaptive and agentic systems, cognition, logic, linguistics, simulation, knowledge management, and more!
- Creative reasoning and general chat performance supplemented with sequelbox/Raiden-DeepSeek-R1

Get the newest SV3: ValiantLabs/Ministral-3-14B-Reasoning-2512-ShiningValiant3

Esper 3.1 is available for Ministral 3 as well: ValiantLabs/Ministral-3-14B-Reasoning-2512-Esper3.1

We're working hard on our next Big New Release, coming out in the next few weeks :)

Help support our releases, donations used for models and datasets: sequelbox/SupportOpenSource

Open source matters. Fight for it with us.

with love and friendship,
allegra

1 reply

reacted to sequelbox's post with 🔥 6 months ago

Post

2101

NEW RELEASE: Esper 3.1 for Ministral 3 14b, 8b, and 3b!

- Esper is our full-stack, full-cycle coding, DevOps, and architecture specialist!
- Our newest, best DeepSeek technical datasets emphasize more challenging queries and tough real-world coding tasks across a variety of programming languages and development paradigms:
- Titanium 3 for coding and reasoning in DevOps and architecture: sequelbox/Titanium3-DeepSeek-V3.1-Terminus
- Tachibana 3 for high-difficulty code production in a variety of topics and programming languages:
- sequelbox/Tachibana3-Part1-DeepSeek-V3.1-Terminus
- sequelbox/Tachibana3-Part2-DeepSeek-V3.2
- Mitakihara for MLOps, AI building, use, expertise, and research: sequelbox/Mitakihara-DeepSeek-R1-0528

Get Esper 3.1 now in all 3 Ministral 3 sizes! (We recommend 14b for general use.)
14b: ValiantLabs/Ministral-3-14B-Reasoning-2512-Esper3.1
8b: ValiantLabs/Ministral-3-8B-Reasoning-2512-Esper3.1
3b: ValiantLabs/Ministral-3-3B-Reasoning-2512-Esper3.1

We'll be bringing more models to Ministral soon, including Shining Valiant 3 :)

We're currently working hard on a big release in a new specialty - hoping to have that up on Valiant Labs before the end of the year! We'll keep pushing the boundaries of what personal-sized AI can do for you.

See our Experimental Reasoning models and open-source datasets: @sequelbox

Help us keep working for open source AI with a donation: sequelbox/SupportOpenSource

with love,
allegra

1 reply

reacted to sequelbox's post with 🤗 7 months ago

Post

2554

NEW RELEASE: UML Generator is here!

- Our newest Experimental Reasoning release: create Unified Modeling Language diagrams to provide analysis and insight into your queries and situations!
- Multi-step reasoning reliably identifies diagram structure before a user response of XMI 2.5.1 code containing the UML diagram. Load the diagram into the UML tool of your choice!
- Trained in a variety of subjects for flexible analysis: software architecture, software development, business processes, systems engineering, data modeling, microservices, reverse engineering and more!

UML Generator available for multiple sizes of gpt-oss and Qwen 3, to provide increased flexibility to the user:
gpt-oss-120b: sequelbox/gpt-oss-120b-UML-Generator
gpt-oss-20b: sequelbox/gpt-oss-20b-UML-Generator
Qwen3-14B: sequelbox/Qwen3-14B-UML-Generator
Qwen3-4B-Thinking-2507: sequelbox/Qwen3-4B-Thinking-2507-UML-Generator

You can also get the UML Generator dataset, to train your own models to use UML Generator Format: sequelbox/UML-Generator-Dataset-DeepSeek-V3.2

Support our experimental open-source research efforts, models and datasets: sequelbox/SupportOpenSource

See our other Experimental Reasoning models: https://huggingface.co/collections/sequelbox/experimental-reasoning-models

with love,
allegra

reacted to sequelbox's post with 👍 8 months ago

Post

1515

NEW RELEASE: Esper 3.1!

- Esper is our full-stack, full-cycle coding, DevOps, and architecture specialist!
- Our newest, best DeepSeek technical datasets emphasize more challenging queries and tough real-world coding tasks across a variety of programming languages and development paradigms:
- Titanium 3 for coding and reasoning in DevOps and architecture: sequelbox/Titanium3-DeepSeek-V3.1-Terminus
- Tachibana 3 for high-difficulty code production in a variety of topics and programming languages:
- sequelbox/Tachibana3-Part1-DeepSeek-V3.1-Terminus
- sequelbox/Tachibana3-Part2-DeepSeek-V3.2
- Mitakihara for MLOps, AI building, use, expertise, and research: sequelbox/Mitakihara-DeepSeek-R1-0528

Our first release in the Esper 3.1 series is built on Qwen3-4B-Thinking-2507. GET IT NOW, FOR EVERYONE: ValiantLabs/Qwen3-4B-Thinking-2507-Esper3.1

We'll be bringing Esper 3.1 to more, larger models as soon as we can; you can help this happen faster with a donation: sequelbox/SupportOpenSource

We're really happy about this one; let us know how Esper 3.1 works for you!

Support open source. It's our only hope for an AI future you'll actually want to live in.

More to come soon!

with our love and appreciation,
allegra

reacted to hba123's post with 🔥🚀 9 months ago

Post

2223

Hey, amazing, awesome people of the beautiful internet 😍🥰

Distillation has been (from my point of view) a main driving factor for the success of hashtag#LLMs - like distilling the knowledge of an amazing big model (say hashtag#DeepSeekv3, or hashtag#GeminiAI) into yours.

Probably, you have done it with minimising a KL divergence, and it somehow worked.

Well, not that well, right?

1️⃣ Your model tends to memorise!
2️⃣ Your model might get the right answer, but its reasoning might be flawed.

To fix those problems, we rethink distillation and process a new approach! A method that is based on constrained RL that comes with nice theoretical guarantees and excellent performance!

Check it out: Rethinking Large Language Model Distillation: A Constrained Markov Decision Process Perspective (2509.22921)

Let us do distillation right! Please upvote if you find it useful!

reacted to giadap's post with 👍❤️ 9 months ago

Post

11145

One of the hardest challenges in AI safety is finding the right balance: how do we protect people from harm without undermining their agency? This tension is especially visible in conversational systems, where safeguards can sometimes feel more paternalistic than supportive.

In my latest piece for Hugging Face, I argue that open source and community-driven approaches offer a promising (though not exclusive) way forward.

✨ Transparency can make safety mechanisms into learning opportunities.
✨ Collaboration with diverse communities makes safeguards more relevant across contexts.
✨ Iteration in the open lets protections evolve rather than freeze into rigid, one-size-fits-all rules.

Of course, this isn’t a silver bullet. Top-down safety measures will still be necessary in some cases. But if we only rely on corporate control, we risk building systems that are safe at the expense of trust and autonomy.

Read the blog post here: https://huggingface.co/blog/giadap/preserving-agency

8 replies

reacted to Kseniase's post with ❤️ 9 months ago

Post

6255

10 awesome advanced LoRA approaches

Low-Rank Adaptation (LoRA) is the go-to method for efficient model fine-tuning that adds small low-rank matrices instead of retraining full models. The field isn’t standing still – new LoRA variants push the limits of efficiency, generalization, and personalization. So we’re sharing 10 of the latest LoRA approaches you should know about:

1. Mixture-of-LoRA-experts → Mixture of Low-Rank Adapter Experts in Generalizable Audio Deepfake Detection (2509.13878)
Adds multiple low-rank adapters (LoRA) into a model’s layers, and a routing mechanism activates the most suitable ones for each input. This lets the model adapt better to new unseen conditions

2. Amortized Bayesian Meta-Learning for LoRA (ABMLL) → Amortized Bayesian Meta-Learning for Low-Rank Adaptation of Large Language Models (2508.14285)
Balances global and task-specific parameters within a Bayesian framework to improve uncertainty calibration and generalization to new tasks without high memory or compute costs

3. AutoLoRA → AutoLoRA: Automatic LoRA Retrieval and Fine-Grained Gated Fusion for Text-to-Image Generation (2508.02107)
Automatically retrieves and dynamically aggregates public LoRAs for stronger T2I generation

4. aLoRA (Activated LoRA) → Activated LoRA: Fine-tuned LLMs for Intrinsics (2504.12397)
Only applies LoRA after invocation, letting the model reuse the base model’s KV cache instead of recomputing the full turn’s KV cache. Efficient in multi-turn conversations

5. LiLoRA (LoRA in LoRA) → LoRA in LoRA: Towards Parameter-Efficient Architecture Expansion for Continual Visual Instruction Tuning (2508.06202)
Shares the LoRA matrix A across tasks and additionally low-rank-decomposes matrix B to cut parameters in continual vision-text MLLMs

6. Sensitivity-LoRA → Sensitivity-LoRA: Low-Load Sensitivity-Based Fine-Tuning for Large Language Models (2509.09119)
Dynamically assigns ranks to weight matrices based on their sensitivity, measured using second-order derivatives

Read further below ↓
Also, subscribe to the Turing Post: https://www.turingpost.com/subscribe

3 replies

reacted to codelion's post with 🔥 10 months ago

Post

5307

I wanted to share a technique that's been working really well for recovering performance after INT4 quantization.

Typically, quantizing the LLM to INT4 (unlike say INT8) for inference can incur some accuracy loss. Instead of accepting the quality loss, we used the FP16 model as a teacher to train a tiny LoRA adapter (rank=16) for the quantized model. The cool part: the model generates its own training data using the Magpie technique so no external datasets needed. This is critical because we want to remain as much as possible in the distribution of the model's natural responses.

Last year Apple's foundational models paper (https://arxiv.org/pdf/2407.21075) had proposed a similar technique and found "By using accuracy-recovery LoRA adapters with only rank 16, Alpaca win rate can be improved by 7-18%, GMS8K accuracy is boosted by 5-10%." (page 47).

We saw similar results on Qwen3-0.6B:

Perplexity: 2.40 → 2.09 (only 5.7% degradation from FP16 baseline)
Memory: Only 0.28GB vs 1.0GB for FP16 (75% reduction)
Speed: 3.0x faster inference than FP16
Quality: Generates correct, optimized code solutions

- Pre-trained adapter: codelion/Qwen3-0.6B-accuracy-recovery-lora
- GitHub repo: https://github.com/codelion/ellora

Happy to answer questions about the implementation or help anyone trying to replicate this. The key insight is that quantization errors are systematic and learnable - a small adapter can bridge the gap without negating the benefits of quantization.

Has anyone else experimented with self-distillation for quantization recovery? Would love to hear about different approaches!

replied to sequelbox's post 10 months ago

I also have a hypothesis that this model can be efficiently downsized not by pruning experts, but by using merges and LoRAs to downsize their unique parameter count. The merge would be most of the shared parameters, and the routing table need not change.

I'm building up a new version of my pipeline to test this hypothesis. I suspect it'd let us get most of the performance in <12B parameters.

replied to sequelbox's post 10 months ago

This is a very cool release! I really enjoy the ShiningValiant series!

Do you see potential to prune experts or layers from the gpt-oss-20b model to downsize it, and then finetune?

sometimesanotion

AI & ML interests

Recent Activity

Organizations

sometimesanotion's activity