Chain-of-Thought Faithfulness of Reasoning Models Varies with Where and How Preference Cues Are Delivered
Paper • 2608.29464 • Published • 15
AI & ML interests
We aim to unify the schema across many different biomedical NLP resources.
Recent Activity
View all activity
Convert dataset to Parquet
1
#8 opened about 1 year ago
by 
SaylorTwift
Dataset Viewer issue: DatasetWithScriptNotSupportedError
2
#11 opened 3 months ago
by 
furkanallen
aryopg
authored a
paper about 1 month ago
Post
2867
SPECK 2 IS ALREADY OUT: specklabs/Speck2-140M
Pretrained on 4x more tokens than the previous releases (20b vs 5b).
Instruct tuned versions are coming soon.
Very interesting models are coming soon too (hint: super long context).
Thanks for everyone supporting!
Pretrained on 4x more tokens than the previous releases (20b vs 5b).
Instruct tuned versions are coming soon.
Very interesting models are coming soon too (hint: super long context).
Thanks for everyone supporting!
- 3 replies
Post
154
NEW SPECK UPDATES:
Just hit #14 and #15 with out FIRST models on Open SLM Leaderboard. The models were trained on 5B tokens, while competing with similarly sized models trained on more than 6-20x the data.
A new base model Speck1.5-140M being trained right now on a higher quality corpus and will be released soon.
SpeckChat3 is coming very soon with 1 million samples, specifically designed to post train small base models.
Also, just to clarify stuff, we will NOT release anything that is NOT MIT licensed EVER. Openness is needed in small language research.
Thanks to everyone supporting the project, and stay tuned for new releases!
Just hit #14 and #15 with out FIRST models on Open SLM Leaderboard. The models were trained on 5B tokens, while competing with similarly sized models trained on more than 6-20x the data.
A new base model Speck1.5-140M being trained right now on a higher quality corpus and will be released soon.
SpeckChat3 is coming very soon with 1 million samples, specifically designed to post train small base models.
Also, just to clarify stuff, we will NOT release anything that is NOT MIT licensed EVER. Openness is needed in small language research.
Thanks to everyone supporting the project, and stay tuned for new releases!
Post
1887
SPECK UPDATES:
1 New instruct model tuned on top of Speck1-140M: specklabs/Speck1-140M-Instruct
2 Instruction tuning datasets
2 GGUFs
Much more coming soon:
Speck1.1-140M-Instruct that is post trained on SpeckChat2 will be coming very soon
New base model Speck1.5-140M is coming with a much higher quality corpus
Thanks to everyone who is already supporting the project, and stay tuned for new releases!
1 New instruct model tuned on top of Speck1-140M: specklabs/Speck1-140M-Instruct
2 Instruction tuning datasets
2 GGUFs
Much more coming soon:
Speck1.1-140M-Instruct that is post trained on SpeckChat2 will be coming very soon
New base model Speck1.5-140M is coming with a much higher quality corpus
Thanks to everyone who is already supporting the project, and stay tuned for new releases!
- 3 replies
Post
2125
FIRST SPECK MODEL RELEASED:
specklabs/Speck1-140M
new models coming very soon (both instruct and much better models), with much much higher training scale as i am getting marenostrum5 access soon!
we will be looking at 100b-2t token budgets :)
specklabs/Speck1-140M
new models coming very soon (both instruct and much better models), with much much higher training scale as i am getting marenostrum5 access soon!
we will be looking at 100b-2t token budgets :)
- 4 replies
Post
2369
Anthropic gave me six months of Claude Max 20x through the Claude for Open Source program, granted based on my Hugging Face work. Thank you 
Anthropic for supporting open source.
So far I've been pointing it at Markdown Minimap, an Obsidian plugin that adds a scrollable IDE-style minimap to your notes. This week I've been clearing a backlog of user-reported issues on it, with Claude often handling them end to end.
https://github.com/Nymbo/Markdown-Minimap — issues and PRs welcome.
So far I've been pointing it at Markdown Minimap, an Obsidian plugin that adds a scrollable IDE-style minimap to your notes. This week I've been clearing a backlog of user-reported issues on it, with Claude often handling them end to end.
https://github.com/Nymbo/Markdown-Minimap — issues and PRs welcome.
Post
200
Hey! I've shared a few synth-like datasets generated in the Synthlabs app using deepseek-v4-flash, qwen3-35B A3B, and Minimax M2.7.
https://huggingface.co/collections/mkurman/synthlabs-datasets
https://huggingface.co/collections/mkurman/synthlabs-datasets
Post
6081
Introducing Inflect-v2, two exceptionally small, open-weight English TTS models at just 3.9M and 9.3M parameters. Both generate speech multiple times faster than real-time on CPU. Despite their size, Inflect-v2 delivers quality that is competitive with much larger lightweight TTS systems, including KittenTTS, Piper, and Supertonic-3.
CPU, CUDA, PyTorch, and ONNX are supported. Apache 2.0.
See it for yourselves:
owensong/Inflect-Micro-v2
owensong/Inflect-Nano-v2
Try the Demos:
Nymbo/Inflect-TTS (unlimited CPU usage)
owensong/Inflect-v2 (ultra-fast ZeroGPU usage)
CPU, CUDA, PyTorch, and ONNX are supported. Apache 2.0.
See it for yourselves:
owensong/Inflect-Micro-v2
owensong/Inflect-Nano-v2
Try the Demos:
Nymbo/Inflect-TTS (unlimited CPU usage)
owensong/Inflect-v2 (ultra-fast ZeroGPU usage)
- 6 replies
albertvillanova
posted an update 3 months ago
Post
4245
🎉 KTO is now part of the stable TRL API
As of Promote KTO to stable API, KTOTrainer and KTOConfig have graduated from trl.experimental to the stable trl API. https://github.com/huggingface/trl/pull/6175
This one closes out a long road. Over the past 6+ months, the "Align KTO with DPO" effort landed ~90 PRs methodically bringing KTO up to the standard we hold for stable trainers, one carefully-scoped change at a time:
- Feature parity with DPO: full VLM support (incl. multi-image), sync_ref_model, PEFT + Liger, ZeRO-3 + PEFT dtype fix, pad_to_multiple_of, activation offloading, IterableDataset and dict eval_dataset, remove_unused_columns, and reference-logprob precomputation at init.
- Consistency with DPO: aligned method order and signatures, tokenization, _prepare_dataset, PEFT handling, ref-model preparation for distributed training, and config layout — plus a new DataCollatorForKTO and output format. Metrics moved into _compute_loss and simplified to direct averages via the shared _metrics attribute.
- Removing legacy baggage: dropped encoder-decoder support, BOS/EOS handling, null_ref_context, generate_during_eval, model_init, preprocess_logits_for_metrics, model/ref adapter names, and several dead config knobs.
- Coverage: a full test suite mirroring DPO, text collator tests, VLM tests, and slow tests.
- The promotion itself: the experimental → stable move (#6175) and shim cleanup (#6287), handled so downstream users get a clean deprecation path.
Honestly, this has been one of the more complex tasks I've taken on since joining the team, not because any single change was hard, but because it demanded sustained consistency across a ~2,000-line trainer, with every branch, comment, and edge case kept in lockstep with DPO.
Huge thanks to everyone who reviewed along the way (especially @qgallouedec ), the incremental review cadence is exactly what kept this maintainable.
KTO now sits on equal footing with our other flagship trainers. 🚀
As of Promote KTO to stable API, KTOTrainer and KTOConfig have graduated from trl.experimental to the stable trl API. https://github.com/huggingface/trl/pull/6175
This one closes out a long road. Over the past 6+ months, the "Align KTO with DPO" effort landed ~90 PRs methodically bringing KTO up to the standard we hold for stable trainers, one carefully-scoped change at a time:
- Feature parity with DPO: full VLM support (incl. multi-image), sync_ref_model, PEFT + Liger, ZeRO-3 + PEFT dtype fix, pad_to_multiple_of, activation offloading, IterableDataset and dict eval_dataset, remove_unused_columns, and reference-logprob precomputation at init.
- Consistency with DPO: aligned method order and signatures, tokenization, _prepare_dataset, PEFT handling, ref-model preparation for distributed training, and config layout — plus a new DataCollatorForKTO and output format. Metrics moved into _compute_loss and simplified to direct averages via the shared _metrics attribute.
- Removing legacy baggage: dropped encoder-decoder support, BOS/EOS handling, null_ref_context, generate_during_eval, model_init, preprocess_logits_for_metrics, model/ref adapter names, and several dead config knobs.
- Coverage: a full test suite mirroring DPO, text collator tests, VLM tests, and slow tests.
- The promotion itself: the experimental → stable move (#6175) and shim cleanup (#6287), handled so downstream users get a clean deprecation path.
Honestly, this has been one of the more complex tasks I've taken on since joining the team, not because any single change was hard, but because it demanded sustained consistency across a ~2,000-line trainer, with every branch, comment, and edge case kept in lockstep with DPO.
Huge thanks to everyone who reviewed along the way (especially @qgallouedec ), the incremental review cadence is exactly what kept this maintainable.
KTO now sits on equal footing with our other flagship trainers. 🚀
- 2 replies
aryopg
authored 2
papers 3 months ago
Post
229
Decades before the modern scaling laws, this paper showed that neural networks behavior under scale follows remarkably predictable laws.
In 1993, researchers at Bell Labs were grappling with a constraint that feels entirely familiar (and contemporary): datasets were outgrowing the available hardware, and training a model to the end was becoming too expensive. To evaluate an architectural tweak to a state-of-the-art model (at the time it was LeNet) on 60,000 samples meant burning up to three weeks of compute time.
To save compute, people would train candidate architectures on small subsets of the data, assuming that the top performer at small scale would remain the top performer at full scale. But with our future wisdom, we know this is not the case.
In "Learning Curves: Asymptotic Values and Rate of Convergence (NeurIPS 93)", using insights from statistical mechanics, they proposed a practical and principled method for predicting the performance of classifiers trained on large datasets (at the time, models were assumed to be large enough). The method was based on a simple power-law modeling of the expected training and test errors.
It is often noted that many of today's breakthroughs in AI and deep learning are actually decades-old concepts that simply lacked the computational power to be tested at the time. While there is some truth to that, it highlights a more valuable lesson: there is immense worth in revisiting early literature and reflecting on foundational ideas we may have prematurely left behind.
So, go explore and find your own inspiration. The current trend has enough champions already!
In 1993, researchers at Bell Labs were grappling with a constraint that feels entirely familiar (and contemporary): datasets were outgrowing the available hardware, and training a model to the end was becoming too expensive. To evaluate an architectural tweak to a state-of-the-art model (at the time it was LeNet) on 60,000 samples meant burning up to three weeks of compute time.
To save compute, people would train candidate architectures on small subsets of the data, assuming that the top performer at small scale would remain the top performer at full scale. But with our future wisdom, we know this is not the case.
In "Learning Curves: Asymptotic Values and Rate of Convergence (NeurIPS 93)", using insights from statistical mechanics, they proposed a practical and principled method for predicting the performance of classifiers trained on large datasets (at the time, models were assumed to be large enough). The method was based on a simple power-law modeling of the expected training and test errors.
It is often noted that many of today's breakthroughs in AI and deep learning are actually decades-old concepts that simply lacked the computational power to be tested at the time. While there is some truth to that, it highlights a more valuable lesson: there is immense worth in revisiting early literature and reflecting on foundational ideas we may have prematurely left behind.
So, go explore and find your own inspiration. The current trend has enough champions already!
- 2 replies
Post
334
It has been more than a decade now since the knowledge distillation paper came out.
Knowledge Distillation (KD) is one of my favorite topics, but I have to confess that I'm not a huge fan of the term because I find it confusing (or at least, it has became so over time).
The idea behind KD is not novel; it was there almost a decade before the paper came out (and arguably even a decade before that, back to 1990-91). But this paper is the one that clicked, the one that made the topic much more popular and introduced it to a broader audience.
First, the timing and the authors played a big role: we have Geoffrey Hinton, Oriol Vinyals, and Jeff Dean here. And second, Geoffrey Hinton is really good at idea branding: Model compression?! No, no, no! Let's call it "Knowledge Distillation" and use evocative terms such as "Dark Knowledge" to describe what is being transferred.
It's a great name, but as time has passed, the term became a bit of a relic. KD is no longer solely about compression (KD used to be introduced as a method for model compression, but now model compression is just one application of KD). And the other thing is that the word "distillation" implies some sort of potency here, that the student is somehow more powerful than the teacher, which is not the case (but many counterarguments could be made, for example, more powerful compared to another model trained with no teacher)
Nevertheless, the paper is incredibly well-written, short, and fun to read. It's one of few papers that I read several times. Check it out, and maybe share your thoughts on the topic with us here!
If you had to choose another name for Knowledge Distillation, what would it be?
Knowledge Distillation (KD) is one of my favorite topics, but I have to confess that I'm not a huge fan of the term because I find it confusing (or at least, it has became so over time).
The idea behind KD is not novel; it was there almost a decade before the paper came out (and arguably even a decade before that, back to 1990-91). But this paper is the one that clicked, the one that made the topic much more popular and introduced it to a broader audience.
First, the timing and the authors played a big role: we have Geoffrey Hinton, Oriol Vinyals, and Jeff Dean here. And second, Geoffrey Hinton is really good at idea branding: Model compression?! No, no, no! Let's call it "Knowledge Distillation" and use evocative terms such as "Dark Knowledge" to describe what is being transferred.
It's a great name, but as time has passed, the term became a bit of a relic. KD is no longer solely about compression (KD used to be introduced as a method for model compression, but now model compression is just one application of KD). And the other thing is that the word "distillation" implies some sort of potency here, that the student is somehow more powerful than the teacher, which is not the case (but many counterarguments could be made, for example, more powerful compared to another model trained with no teacher)
Nevertheless, the paper is incredibly well-written, short, and fun to read. It's one of few papers that I read several times. Check it out, and maybe share your thoughts on the topic with us here!
If you had to choose another name for Knowledge Distillation, what would it be?
- 6 replies
Post
2696
What if you could train a model on just 10 images instead of 60,000 and still get close to the same performance?
Traditional machine learning requires thousands, even millions, of data points to achieve high accuracy. But what if we could "distill" the entire dataset into just a few synthetic samples?
This is what Dataset Distillation offers. Unlike traditional knowledge distillation, we keep the model fixed and distill the knowledge contained in a massive training set into a tiny set of synthetic distilled images.
The goal is to train a model on this ultra-small set and achieve performance that almost matches what the same model would get when trained on the massive original dataset.
For example, training on only 10 distilled MNIST images (this is equivalent to a single image per class) yields 94% accuracy, compared to 99% when training on the full 60,000 images.
Interestingly, these distilled images look significantly different (as you can see in the image below) from natural images because they are optimized for model training rather than for matching the correct data distribution.
But that's not all.
Most importantly, this same method opens the door to a potent form of data poisoning. Because distilled images are specifically optimized for rapid learning, an attacker can create a tiny set of adversarial distilled images to cause a well-trained model to forget or misclassify a specific category.
What I find fascinating about dataset distillation is this: it mimics human-like learning by letting a model grasp a concept from a single example, but it does so using alien synthetic images that mean absolutely nothing to a human eye!
What about you? What are your thoughts on it?
Traditional machine learning requires thousands, even millions, of data points to achieve high accuracy. But what if we could "distill" the entire dataset into just a few synthetic samples?
This is what Dataset Distillation offers. Unlike traditional knowledge distillation, we keep the model fixed and distill the knowledge contained in a massive training set into a tiny set of synthetic distilled images.
The goal is to train a model on this ultra-small set and achieve performance that almost matches what the same model would get when trained on the massive original dataset.
For example, training on only 10 distilled MNIST images (this is equivalent to a single image per class) yields 94% accuracy, compared to 99% when training on the full 60,000 images.
Interestingly, these distilled images look significantly different (as you can see in the image below) from natural images because they are optimized for model training rather than for matching the correct data distribution.
But that's not all.
Most importantly, this same method opens the door to a potent form of data poisoning. Because distilled images are specifically optimized for rapid learning, an attacker can create a tiny set of adversarial distilled images to cause a well-trained model to forget or misclassify a specific category.
What I find fascinating about dataset distillation is this: it mimics human-like learning by letting a model grasp a concept from a single example, but it does so using alien synthetic images that mean absolutely nothing to a human eye!
What about you? What are your thoughts on it?
- 2 replies
WojciechKusa
authored 2
papers 4 months ago
Post
5013
It was supposed to be a failed experiment. Instead, it led to the discovery of one of the most intriguing phenomena in neural networks, simply because a researcher forgot to turn it off and left it running....for a week!
In 2022, researchers at OpenAI were studying how neural networks generalize from their training data. For this task, they were training small transformer models to perform modular arithmetic.
The thing is, neural networks are weird. When a model has an abundance of parameters (like neural nets), it can easily overfit. It essentially memorizes its training data, scoring a perfect 100% accuracy when tested on it, but remains completely clueless when faced with any new instances not present in the training set (close to 0 accuracy). It is like memorizing 1 + 2 = 3 without understanding the concept of addition, so if 2 + 3 wasn't in the training set, the model fails miserably!
Usually, when a model overfits like this, people just cut their losses, turn off the experiment, and move on with their lives.
But sometimes they forget. And that is exactly what happened to our researchers at OpenAI. A week later, they checked back in, and a miracle had happened!
They discovered Grokking (And no, this has nothing to do with xAI's Grok , the term was originally coined by sci-fi author Robert Heinlein to mean understanding something so deeply that it becomes part of you). Grokking is when a neural network suddenly and abruptly learns to generalize long after it has overfitted. Just take a look at the graph in the image below!
Spooky, right! I told you neural nets are weird!
In 2022, researchers at OpenAI were studying how neural networks generalize from their training data. For this task, they were training small transformer models to perform modular arithmetic.
The thing is, neural networks are weird. When a model has an abundance of parameters (like neural nets), it can easily overfit. It essentially memorizes its training data, scoring a perfect 100% accuracy when tested on it, but remains completely clueless when faced with any new instances not present in the training set (close to 0 accuracy). It is like memorizing 1 + 2 = 3 without understanding the concept of addition, so if 2 + 3 wasn't in the training set, the model fails miserably!
Usually, when a model overfits like this, people just cut their losses, turn off the experiment, and move on with their lives.
But sometimes they forget. And that is exactly what happened to our researchers at OpenAI. A week later, they checked back in, and a miracle had happened!
They discovered Grokking (And no, this has nothing to do with xAI's Grok , the term was originally coined by sci-fi author Robert Heinlein to mean understanding something so deeply that it becomes part of you). Grokking is when a neural network suddenly and abruptly learns to generalize long after it has overfitted. Just take a look at the graph in the image below!
Spooky, right! I told you neural nets are weird!
- 4 replies