Distillation-Hugs (Distillation Hugs)

andito

posted an update 4 days ago

Post

2352

Extremely bullish on @CohereForAI 's Aya Vision (8B & 32B) - new SOTA open-weight VLMs

- 8B wins up to 81% of the time in its class, better than Gemini Flash
- 32B beats Llama 3.2 90B!
- Covers 23 languages, excels in image captioning, VQA & more
- Integrated on transformers from Day 0!

Efficient multimodal models are here to stay!!🔥
Check out their blog! https://huggingface.co./blog/aya-vision

tomaarsen

authored a paper 16 days ago

MMTEB: Massive Multilingual Text Embedding Benchmark

Paper • 2502.13595 • Published 18 days ago • 31

lewtun

posted an update 27 days ago

Post

4787

Introducing OpenR1-Math-220k!

open-r1/OpenR1-Math-220k

The community has been busy distilling DeepSeek-R1 from inference providers, but we decided to have a go at doing it ourselves from scratch 💪

What’s new compared to existing reasoning datasets?

♾ Based on AI-MO/NuminaMath-1.5: we focus on math reasoning traces and generate answers for problems in NuminaMath 1.5, an improved version of the popular NuminaMath-CoT dataset.

🐳 800k R1 reasoning traces: We generate two answers for 400k problems using DeepSeek R1. The filtered dataset contains 220k problems with correct reasoning traces.

📀 512 H100s running locally: Instead of relying on an API, we leverage vLLM and SGLang to run generations locally on our science cluster, generating 180k reasoning traces per day.

⏳ Automated filtering: We apply Math Verify to only retain problems with at least one correct answer. We also leverage Llama3.3-70B-Instruct as a judge to retrieve more correct examples (e.g for cases with malformed answers that can’t be verified with a rules-based parser)

📊 We match the performance of DeepSeek-Distill-Qwen-7B by finetuning Qwen-7B-Math-Instruct on our dataset.

🔎 Read our blog post for all the nitty gritty details: https://huggingface.co./blog/open-r1/update-2

andito

authored a paper about 1 month ago

SmolLM2: When Smol Goes Big -- Data-Centric Training of a Small Language Model

Paper • 2502.02737 • Published Feb 4 • 199

lewtun

authored a paper about 1 month ago

SmolLM2: When Smol Goes Big -- Data-Centric Training of a Small Language Model

Paper • 2502.02737 • Published Feb 4 • 199

lewtun

posted an update about 1 month ago

Post

10223

We are reproducing the full DeepSeek R1 data and training pipeline so everybody can use their recipe. Instead of doing it in secret we can do it together in the open!

🧪 Step 1: replicate the R1-Distill models by distilling a high-quality reasoning corpus from DeepSeek-R1.

🧠 Step 2: replicate the pure RL pipeline that DeepSeek used to create R1-Zero. This will involve curating new, large-scale datasets for math, reasoning, and code.

🔥 Step 3: show we can go from base model -> SFT -> RL via multi-stage training.

Follow along: https://github.com/huggingface/open-r1

5 replies

·

tomaarsen

posted an update about 2 months ago

Post

2193

I just released Sentence Transformers v3.4.0, featuring a memory leak fix, compatibility between the powerful Cached... losses and the Matryoshka loss modifier, and a bunch of fixes & small features.

🪆 Matryoshka & Cached loss compatibility
It is now possible to combine the powerful Cached... losses (which use in-batch negatives & a caching mechanism to allow for endless batch size & negatives) with the Matryoshka loss modifier which modifies a base loss such that it is trained not only on the maximum dimensionality (e.g. 1024 dimensions), but also on many lower dimensions (e.g. 768, 512, 256, 128, 64, 32).
After training, these models' embeddings can be truncated for faster retrieval, etc.

🎞️ Resolve memory leak when Model and Trainer are reinitialized
Due to a circular dependency between Trainer -> Model -> ModelCardData -> Trainer, deleting both the trainer & model still didn't free up the memory.
This led to a memory leak in scripts where you repeatedly do so.

➕ New Features
Many new small features, e.g. multi-GPU support for 'mine_hard_negatives', a 'margin' parameter to TripletEvaluator, and Matthews Correlation Coefficient in the BinaryClassificationEvaluator.

🐛 Bug Fixes
Also a bunch of fixes, for example that subsequent batches were not sorted when using the "no_duplicates" batch sampler. See the release notes for more details.

Full release notes: https://github.com/UKPLab/sentence-transformers/releases/tag/v3.4.0

Big thanks to all community members who assisted in this release. 10 folks with their first contribution this time around!

andito

posted an update about 2 months ago

Post

1630

𝗜𝗻𝘁𝗿𝗼𝗱𝘂𝗰𝗶𝗻𝗴 𝘁𝗵𝗲 𝘄𝗼𝗿𝗹𝗱'𝘀 𝘀𝗺𝗮𝗹𝗹𝗲𝘀𝘁 𝘃𝗶𝘀𝗶𝗼𝗻 𝗹𝗮𝗻𝗴𝘂𝗮𝗴𝗲 𝗺𝗼𝗱𝗲𝗹!

We’re thrilled to share 𝗦𝗺𝗼𝗹𝗩𝗟𝗠 (256M & 500M)—the smallest Visual Language Models ever built. Think: running on <1GB of GPU memory—you can fine-tune it on your laptop and run it on your toaster!

Why It’s Game-Changing:
- 𝗢𝘂𝘁𝗽𝗲𝗿𝗳𝗼𝗿𝗺𝘀 𝗟𝗮𝗿𝗴𝗲𝗿 𝗠𝗼𝗱𝗲𝗹𝘀: Even the 256M model surpasses our SOTA 80B-parameter model from just 17 months ago. Over 300x reduction!
𝗠𝗶𝗴𝗵𝘁𝘆 𝗘𝗳𝗳𝗶𝗰𝗶𝗲𝗻𝗰𝘆: The 256M version delivers 80% of our 2.2B model’s performance, and the 500M version hits 90%
𝗟𝗶𝗴𝗵𝘁𝗻𝗶𝗻𝗴-𝗙𝗮𝘀𝘁 𝗦𝗲𝗮𝗿𝗰𝗵: SmolVLM integrates with ColiPali for state-of-the-art retrieval speeds—on par with models 10x bigger. That means cheaper, faster indexing and real-world impact.

What’s New Under the Hood:
- 𝗡𝗲𝘄 𝗩𝗶𝘀𝗶𝗼𝗻 𝗘𝗻𝗰𝗼𝗱𝗲𝗿: Smaller overall size (400M -> 93M), but with higher resolution.
- 𝗛𝗶𝗴𝗵𝗲𝗿 𝗣𝗶𝘅𝗲𝗹𝘀/𝗧𝗼𝗸𝗲𝗻: 4096 vs. 1820—more efficient image processing.
- 𝗦𝗺𝗮𝗿𝘁 𝗧𝗼𝗸𝗲𝗻𝗶𝘇𝗮𝘁𝗶𝗼𝗻: Faster training and a performance boost.

Check our blog: https://huggingface.co./blog/smolervlm
The models: HuggingFaceTB/smolvlm-256m-and-500m-6791fafc5bb0ab8acc960fb0
The demo: HuggingFaceTB/SmolVLM-256M-Demo

1 reply

·

tomaarsen

posted an update about 2 months ago

Post

4649

🏎️ Today I'm introducing a method to train static embedding models that run 100x to 400x faster on CPU than common embedding models, while retaining 85%+ of the quality! Including 2 fully open models: training scripts, datasets, metrics.

We apply our recipe to train 2 Static Embedding models that we release today! We release:
2️⃣ an English Retrieval model and a general-purpose Multilingual similarity model (e.g. classification, clustering, etc.), both Apache 2.0
🧠 my modern training strategy: ideation -> dataset choice -> implementation -> evaluation
📜 my training scripts, using the Sentence Transformers library
📊 my Weights & Biases reports with losses & metrics
📕 my list of 30 training and 13 evaluation datasets

The 2 Static Embedding models have the following properties:
🏎️ Extremely fast, e.g. 107500 sentences per second on a consumer CPU, compared to 270 for 'all-mpnet-base-v2' and 56 for 'gte-large-en-v1.5'
0️⃣ Zero active parameters: No Transformer blocks, no attention, not even a matrix multiplication. Super speed!
📏 No maximum sequence length! Embed texts at any length (note: longer texts may embed worse)
📐 Linear instead of exponential complexity: 2x longer text takes 2x longer, instead of 2.5x or more.
🪆 Matryoshka support: allow you to truncate embeddings with minimal performance loss (e.g. 4x smaller with a 0.56% perf. decrease for English Similarity tasks)

Check out the full blogpost if you'd like to 1) use these lightning-fast models or 2) learn how to train them with consumer-level hardware: https://huggingface.co./blog/static-embeddings

The blogpost contains a lengthy list of possible advancements; I'm very confident that our 2 models are only the tip of the iceberg, and we may be able to get even better performance.

Alternatively, check out the models:
* sentence-transformers/static-retrieval-mrl-en-v1
* sentence-transformers/static-similarity-mrl-multilingual-v1

1 reply

·

lewtun

posted an update 2 months ago

Post

3881

I was initially pretty sceptical about Meta's Coconut paper [1] because the largest perf gains were reported on toy linguistic problems. However, these results on machine translation are pretty impressive!

https://x.com/casper_hansen_/status/1875872309996855343

Together with the recent PRIME method [2] for scaling RL, reasoning for open models is looking pretty exciting for 2025!

[1] Training Large Language Models to Reason in a Continuous Latent Space (2412.06769)
[2] https://huggingface.co./blog/ganqu/prime

tomaarsen

posted an update 2 months ago

Post

3032

That didn't take long! Nomic AI has finetuned the new ModernBERT-base encoder model into a strong embedding model for search, classification, clustering and more!

Details:
🤖 Based on ModernBERT-base with 149M parameters.
📊 Outperforms both nomic-embed-text-v1 and nomic-embed-text-v1.5 on MTEB!
🏎️ Immediate FA2 and unpacking support for super efficient inference.
🪆 Trained with Matryoshka support, i.e. 2 valid output dimensionalities: 768 and 256.
➡️ Maximum sequence length of 8192 tokens!
2️⃣ Trained in 2 stages: unsupervised contrastive data -> high quality labeled datasets.
➕ Integrated in Sentence Transformers, Transformers, LangChain, LlamaIndex, Haystack, etc.
🏛️ Apache 2.0 licensed: fully commercially permissible

Try it out here: nomic-ai/modernbert-embed-base

Very nice work by Zach Nussbaum and colleagues at Nomic AI.

lewtun

posted an update 2 months ago

Post

2308

This paper ( HuatuoGPT-o1, Towards Medical Complex Reasoning with LLMs (2412.18925)) has a really interesting recipe for inducing o1-like behaviour in Llama models:

* Iteratively sample CoTs from the model, using a mix of different search strategies. This gives you something like Stream of Search via prompting.
* Verify correctness of each CoT using GPT-4o (needed because exact match doesn't work well in medicine where there are lots of aliases)
* Use GPT-4o to reformat the concatenated CoTs into a single stream that includes smooth transitions like "hmm, wait" etc that one sees in o1
* Use the resulting data for SFT & RL
* Use sparse rewards from GPT-4o to guide RL training. They find RL gives an average ~3 point boost across medical benchmarks and SFT on this data already gives a strong improvement.

Applying this strategy to other domains could be quite promising, provided the training data can be formulated with verifiable problems!

1 reply

·

tomaarsen

authored a paper 3 months ago

Smarter, Better, Faster, Longer: A Modern Bidirectional Encoder for Fast, Memory Efficient, and Long Context Finetuning and Inference

Paper • 2412.13663 • Published Dec 18, 2024 • 134

lewtun

posted an update 3 months ago

Post

6928

We outperform Llama 70B with Llama 3B on hard math by scaling test-time compute 🔥

How? By combining step-wise reward models with tree search algorithms :)

We show that smol models can match or exceed the performance of their much larger siblings when given enough "time to think"

We're open sourcing the full recipe and sharing a detailed blog post.

In our blog post we cover:

📈 Compute-optimal scaling: How we implemented DeepMind's recipe to boost the mathematical capabilities of open models at test-time.

🎄 Diverse Verifier Tree Search (DVTS): An unpublished extension we developed to the verifier-guided tree search technique. This simple yet effective method improves diversity and delivers better performance, particularly at large test-time compute budgets.

🧭 Search and Learn: A lightweight toolkit for implementing search strategies with LLMs and built for speed with vLLM

Here's the links:

- Blog post: HuggingFaceH4/blogpost-scaling-test-time-compute

- Code: https://github.com/huggingface/search-and-learn

Enjoy!

2 replies

·

andito

posted an update 3 months ago

Post

1962

SmolVLM speeding locally on a laptop thanks to mlx-vlm and
@Gradio ! Try it with two lines:
pip install git+https://github.com/andimarafioti/mlx-vlm.git@stream-generate-fix
python -m mlx_vlm.chat_ui --model mlx-community/SmolVLM-Instruct-8bit

Gotta love the MLX community! Big thanks to @pcuenq and @prince_canuma !

andito

posted an update 3 months ago

Post

3370

Let's go! We are releasing SmolVLM, a smol 2B VLM built for on-device inference that outperforms all models at similar GPU RAM usage and tokens throughputs.

- SmolVLM generates tokens 7.5 to 16 times faster than Qwen2-VL! 🤯
- Other models at this size crash a laptop, but SmolVLM comfortably generates 17 tokens/sec on a macbook! 🚀
- SmolVLM can be fine-tuned on a Google collab! Or process millions of documents with a consumer GPU!
- SmolVLM even outperforms larger models in video benchmarks, despite not even being trained on videos!

Check out more!
Demo: HuggingFaceTB/SmolVLM
Blog: https://huggingface.co./blog/smolvlm
Model: HuggingFaceTB/SmolVLM-Instruct
Fine-tuning script: https://github.com/huggingface/smollm/blob/main/finetuning/Smol_VLM_FT.ipynb

tomaarsen

posted an update 4 months ago

Post

5766

I just released Sentence Transformers v3.3.0 & it's huge! 4.5x speedup for CPU with OpenVINO int8 static quantization, training with prompts for a free perf. boost, PEFT integration, evaluation on NanoBEIR, and more! Details:

1. We integrate Post-Training Static Quantization using OpenVINO, a very efficient solution for CPUs that processes 4.78x as many texts per second on average, while only hurting performance by 0.36% on average. There's a new export_static_quantized_openvino_model method to quantize a model.

2. We add the option to train with prompts, e.g. strings like "query: ", "search_document: " or "Represent this sentence for searching relevant passages: ". It's as simple as using the prompts argument in SentenceTransformerTrainingArguments. Our experiments show that you can easily reach 0.66% to 0.90% relative performance improvement on NDCG@10 at no extra cost by adding "query: " before each training query and "document: " before each training answer.

3. Sentence Transformers now supports training PEFT adapters via 7 new methods for adding new adapters or loading pre-trained ones. You can also directly load a trained adapter with SentenceTransformer as if it's a normal model. Very useful for e.g. 1) training multiple adapters on 1 base model, 2) training bigger models than otherwise possible, or 3) cheaply hosting multiple models by switching multiple adapters on 1 base model.

4. We added easy evaluation on NanoBEIR, a subset of BEIR a.k.a. the MTEB Retrieval benchmark. It contains 13 datasets with 50 queries and up to 10k documents each. Evaluation is fast, and can easily be done during training to track your model's performance on general-purpose information retrieval tasks.

Additionally, we also deprecate Python 3.8, add better compatibility with Transformers v4.46.0, and more. Read the full release notes here: https://github.com/UKPLab/sentence-transformers/releases/tag/v3.3.0

tomaarsen

posted an update 5 months ago

Post

7090

📣 Sentence Transformers v3.2.0 is out, marking the biggest release for inference in 2 years! 2 new backends for embedding models: ONNX (+ optimization & quantization) and OpenVINO, allowing for speedups up to 2x-3x AND Static Embeddings for 500x speedups at 10-20% accuracy cost.

1️⃣ ONNX Backend: This backend uses the ONNX Runtime to accelerate model inference on both CPU and GPU, reaching up to 1.4x-3x speedup depending on the precision. We also introduce 2 helper methods for optimizing and quantizing models for (much) faster inference.
2️⃣ OpenVINO Backend: This backend uses Intel their OpenVINO instead, outperforming ONNX in some situations on CPU.

Usage is as simple as SentenceTransformer("all-MiniLM-L6-v2", backend="onnx"). Does your model not have an ONNX or OpenVINO file yet? No worries - it'll be autoexported for you. Thank me later 😉

🔒 Another major new feature is Static Embeddings: think word embeddings like GLoVe and word2vec, but modernized. Static Embeddings are bags of token embeddings that are summed together to create text embeddings, allowing for lightning-fast embeddings that don't require any neural networks. They're initialized in one of 2 ways:

1️⃣ via Model2Vec, a new technique for distilling any Sentence Transformer models into static embeddings. Either via a pre-distilled model with from_model2vec or with from_distillation where you do the distillation yourself. It'll only take 5 seconds on GPU & 2 minutes on CPU, no dataset needed.
2️⃣ Random initialization. This requires finetuning, but finetuning is extremely quick (e.g. I trained with 3 million pairs in 7 minutes). My final model was 6.6% worse than bge-base-en-v1.5, but 500x faster on CPU.

Full release notes: https://github.com/UKPLab/sentence-transformers/releases/tag/v3.2.0
Documentation on Speeding up Inference: https://sbert.net/docs/sentence_transformer/usage/efficiency.html

1 reply

·

tomaarsen

posted an update 6 months ago

Post

2048

I've just shipped the Sentence Transformers v3.1.1 patch release, fixing the hard negatives mining utility for some models. This utility is extremely useful to get more performance out of your embedding training data.

⛏ Hard negatives are texts that are rather similar to some anchor text (e.g. a query), but are not the correct match. They're difficult for a model to distinguish from the correct answer, often resulting in a stronger model after training.
mine_hard_negatives docs: https://sbert.net/docs/package_reference/util.html#sentence_transformers.util.mine_hard_negatives

🔓 Beyond that, this release removes the numpy<2 restriction from v3.1.0. This was previously required for Windows as not all third-party libraries were updated to support numpy v2. With Sentence Transformers, you can now choose v1 or v2 of numpy.

Check out the full release notes here: https://github.com/UKPLab/sentence-transformers/releases/tag/v3.1.1

I'm looking forward to releasing v3.2, I have some exciting things planned 🚀

tomaarsen

posted an update 6 months ago

Post

2089

🎉SetFit v1.1.0 is out! Training efficient classifiers on CPU or GPU now uses the Sentence Transformers Trainer, and we resolved a lot of issues caused by updates of third-party libraries (like Transformers). Details:

Training a SetFit classifier model consists of 2 phases:
1. Finetuning a Sentence Transformer embedding model
2. Training a Classifier to map embeddings -> classes

🔌The first phase now uses the SentenceTransformerTrainer that was introduced in the Sentence Transformers v3 update. This brings some immediate upsides like MultiGPU support, without any (intended) breaking changes.

➡️ Beyond that, we softly deprecated the "evaluation_strategy" argument in favor of "eval_strategy" (following a Transformers deprecation), and deprecated Python 3.7. In return, we add official support for Python 3.11 and 3.12.

✨ There's some more minor changes too, like max_steps and eval_max_steps now being a hard limit instead of an approximate one, training/validation losses now logging nicely in Notebooks, and the "device" parameter no longer being ignored in some situations.

Check out the full release notes here: https://github.com/huggingface/setfit/releases/tag/v1.1.0
Or read the documentation: https://huggingface.co./docs/setfit
Or check out the public SetFit models for inspiration: https://huggingface.co./models?library=setfit&sort=created

P.s. the model in the code snippet trained in 1 minute and it can classify ~6000 sentences per second on my GPU.

Distillation Hugs

AI & ML interests

Recent Activity

Distillation-Hugs's activity

MMTEB: Massive Multilingual Text Embedding Benchmark

SmolLM2: When Smol Goes Big -- Data-Centric Training of a Small Language Model

SmolLM2: When Smol Goes Big -- Data-Centric Training of a Small Language Model

Smarter, Better, Faster, Longer: A Modern Bidirectional Encoder for Fast, Memory Efficient, and Long Context Finetuning and Inference

AI & ML interests

Recent Activity

Team members 11

Distillation-Hugs's activity