Post
1468
Lacking vllm support for Transformers v5, frustrating only me?
Akhil Theerthala PRO
Akhil-Theerthala
AI & ML interests
None yet
Recent Activity
liked
a dataset
25 minutes ago
birdsql/bird_sql_dev_20251106
reacted
to
mahimairaja's
post
with ๐
4 days ago
Lacking vllm support for Transformers v5, frustrating only me?
posted
an
update
13 days ago
I'm thrilled to share our new paper, "FinForge: Semi-Synthetic Financial Benchmark Generation" in AI4Finance, AAAI'26!
Key Contributions:
- FinForge Framework: A hybrid pipeline integrating manual/programmatic corpus construction with rigorous LM-based synthesis.
- FinForge-5k Dataset: A new snapshot benchmark comprising over 5,000 human-validated Q&A pairs across 11 financial subdomains, derived from a curated corpus of 100,000 verified documents (143M tokens).
- Benchmarking Results: Evaluation of state-of-the-art open and closed-source models reveals significant variance in financial reasoning capabilities, with leading models achieving approximately 80% accuracy.
Huge thanks to my co-authors @glennmatlin , Anant Gupta, Anirudh JM, Rayan Castilla, and Yi Mei Ng for this collaboration.
You can read the paper: https://huggingface.co/papers/2601.06747
Organizations
reacted to
mahimairaja's
post with ๐
4 days ago
posted
an
update
13 days ago
Post
326
I'm thrilled to share our new paper, "FinForge: Semi-Synthetic Financial Benchmark Generation" in AI4Finance, AAAI'26!
Key Contributions:
- FinForge Framework: A hybrid pipeline integrating manual/programmatic corpus construction with rigorous LM-based synthesis.
- FinForge-5k Dataset: A new snapshot benchmark comprising over 5,000 human-validated Q&A pairs across 11 financial subdomains, derived from a curated corpus of 100,000 verified documents (143M tokens).
- Benchmarking Results: Evaluation of state-of-the-art open and closed-source models reveals significant variance in financial reasoning capabilities, with leading models achieving approximately 80% accuracy.
Huge thanks to my co-authors @glennmatlin , Anant Gupta, Anirudh JM, Rayan Castilla, and Yi Mei Ng for this collaboration.
You can read the paper: FinForge: Semi-Synthetic Financial Benchmark Generation (2601.06747)
Key Contributions:
- FinForge Framework: A hybrid pipeline integrating manual/programmatic corpus construction with rigorous LM-based synthesis.
- FinForge-5k Dataset: A new snapshot benchmark comprising over 5,000 human-validated Q&A pairs across 11 financial subdomains, derived from a curated corpus of 100,000 verified documents (143M tokens).
- Benchmarking Results: Evaluation of state-of-the-art open and closed-source models reveals significant variance in financial reasoning capabilities, with leading models achieving approximately 80% accuracy.
Huge thanks to my co-authors @glennmatlin , Anant Gupta, Anirudh JM, Rayan Castilla, and Yi Mei Ng for this collaboration.
You can read the paper: FinForge: Semi-Synthetic Financial Benchmark Generation (2601.06747)
reacted to
tsungyi's
post with ๐ฅ
18 days ago
Post
2635
๐ Exciting News โ NVIDIA Cosmos is celebrating its 1st birthday and has hit 5 MILLION downloads! ๐
In just one year, the Cosmos ecosystem has grown rapidly:
๐ง Cosmos Reason and Cosmos Predict have surpassed 2 MILLION downloads each on @HuggingFace , topping physical AI leaderboards
๐ Cosmos Transfer is enabling adaptation across domains and tasks
๐ฎ Cosmos Cookbook is the go-to hub for recipes from developers and partners like Uber and IntBot.
Thank you to our amazing developer community for making this possible. Here's to pushing the boundaries of world foundation models together!
๐ง๐ปโ๐ณRead the Cosmos Cookbook: https://nvda.ws/4qevli8
๐ Explore Models & Datasets: https://huggingface.co/collections/nvidia/nvidia-cosmos-2
In just one year, the Cosmos ecosystem has grown rapidly:
๐ง Cosmos Reason and Cosmos Predict have surpassed 2 MILLION downloads each on @HuggingFace , topping physical AI leaderboards
๐ Cosmos Transfer is enabling adaptation across domains and tasks
๐ฎ Cosmos Cookbook is the go-to hub for recipes from developers and partners like Uber and IntBot.
Thank you to our amazing developer community for making this possible. Here's to pushing the boundaries of world foundation models together!
๐ง๐ปโ๐ณRead the Cosmos Cookbook: https://nvda.ws/4qevli8
๐ Explore Models & Datasets: https://huggingface.co/collections/nvidia/nvidia-cosmos-2
reacted to
their
post with ๐
18 days ago
Post
418
Is it better to show a model too many images once (Diversity), or extract as much information from a small set of images?
I have always wanted to do an ablation study on this and recently I got the chance to do exactly that. Why? In applied domains like robotics, manufacturing, or banking, we rarely have the luxury of internet-scale diverse image datasets. We are often "Data Poor" in terms of diversity but "Data Rich" in depth.
The takeaway? Density is efficient for facts but dangerous for reasoning (logical collapse) if you don't have larger scale data.
More details:
https://huggingface.co/blog/Akhil-Theerthala/diversity-density-for-vision-language-models
I have always wanted to do an ablation study on this and recently I got the chance to do exactly that. Why? In applied domains like robotics, manufacturing, or banking, we rarely have the luxury of internet-scale diverse image datasets. We are often "Data Poor" in terms of diversity but "Data Rich" in depth.
The takeaway? Density is efficient for facts but dangerous for reasoning (logical collapse) if you don't have larger scale data.
More details:
https://huggingface.co/blog/Akhil-Theerthala/diversity-density-for-vision-language-models
-
- 1 reply
replied to
their
post
20 days ago
The collection of artefacts for the blogpost can be found here
https://hf.co/collections/Akhil-Theerthala/density-vs-diversity-blogpost
posted
an
update
21 days ago
Post
418
Is it better to show a model too many images once (Diversity), or extract as much information from a small set of images?
I have always wanted to do an ablation study on this and recently I got the chance to do exactly that. Why? In applied domains like robotics, manufacturing, or banking, we rarely have the luxury of internet-scale diverse image datasets. We are often "Data Poor" in terms of diversity but "Data Rich" in depth.
The takeaway? Density is efficient for facts but dangerous for reasoning (logical collapse) if you don't have larger scale data.
More details:
https://huggingface.co/blog/Akhil-Theerthala/diversity-density-for-vision-language-models
I have always wanted to do an ablation study on this and recently I got the chance to do exactly that. Why? In applied domains like robotics, manufacturing, or banking, we rarely have the luxury of internet-scale diverse image datasets. We are often "Data Poor" in terms of diversity but "Data Rich" in depth.
The takeaway? Density is efficient for facts but dangerous for reasoning (logical collapse) if you don't have larger scale data.
More details:
https://huggingface.co/blog/Akhil-Theerthala/diversity-density-for-vision-language-models
-
- 1 reply
reacted to
omarkamali's
post with ๐
about 2 months ago
Post
5226
Exciting updates to the Wikipedia Monthly dataset for November! ๐
ใป Fixed a bug to remove infobox leftovers and other wiki markers such as
ใป New python package https://pypi.org/project/wikisets: a dataset builder with efficient sampling so you can combine the languages you want seamlessly for any date (ideal for pretraining data but works for any purpose)
ใป Moved the pipeline to a large server. Much higher costs but with better reliability and predictability (let me know if you'd like to sponsor this!).
ใป Dataset sizes are unfortunately missing for this month due to shenanigans with the migration, but should be back in December's update.
Check out the dataset:
omarkamali/wikipedia-monthly
ใป Fixed a bug to remove infobox leftovers and other wiki markers such as
__TOC__ใป New python package https://pypi.org/project/wikisets: a dataset builder with efficient sampling so you can combine the languages you want seamlessly for any date (ideal for pretraining data but works for any purpose)
ใป Moved the pipeline to a large server. Much higher costs but with better reliability and predictability (let me know if you'd like to sponsor this!).
ใป Dataset sizes are unfortunately missing for this month due to shenanigans with the migration, but should be back in December's update.
Check out the dataset:
omarkamali/wikipedia-monthly
posted
an
update
2 months ago
Post
1355
For anyone who likes to get their papers summarized, I've tried to make a bentocard summarizer for research papers, that enables us to get overviews of the papers, with possible deepdives where interested. Of course, this has to be improved a lot, so do check the space and give your feedback!
Akhil-Theerthala/PaperStack
Some sample bentogrids are as follows:
Akhil-Theerthala/PaperStack
Some sample bentogrids are as follows:
reacted to
codelion's
post with ๐ฅ
5 months ago
Post
4958
I recently added a recipe in ellora to improve reasoning capabilities to Gemma-3-1B using self-supervised learning. Model now shows step-by-step thinking in <think> tags before answering.
Logic puzzle accuracy: 61% โ 84%. 3 hours training on single GPU. ๐ง
Used GRPO where model generates multiple responses and learns to prefer better reasoning. Works surprisingly well for making smaller models more transparent.
๐ Colab: https://colab.research.google.com/github/codelion/ellora/blob/main/Ellora_Recipe_2_Reasoning_LoRA_with_Self-Rewarding_GRPO.ipynb
๐ค Model: codelion/gemma-3-1b-it-reasoning-grpo-lora
๐ป Code: https://github.com/codelion/ellora
Logic puzzle accuracy: 61% โ 84%. 3 hours training on single GPU. ๐ง
Used GRPO where model generates multiple responses and learns to prefer better reasoning. Works surprisingly well for making smaller models more transparent.
๐ Colab: https://colab.research.google.com/github/codelion/ellora/blob/main/Ellora_Recipe_2_Reasoning_LoRA_with_Self-Rewarding_GRPO.ipynb
๐ค Model: codelion/gemma-3-1b-it-reasoning-grpo-lora
๐ป Code: https://github.com/codelion/ellora
- 1 reply
replied to
their
post
6 months ago
Sure! Would love to have a chat!
posted
an
update
6 months ago
Post
2250
I'm excited to announce that I've just released the newest versions of my Kuvera models and the expanded Personal Finance Reasoning dataset on Hugging Face!
What's new:
I've expanded the Personal Finance Reasoning Dataset, which now includes 18.9k samples of real-world financial questions paired with detailed, empathetic answers. The previous generation pipeline was also streamlined with better psychological context and response validations.
I've also released new Kuvera models trained on this improved dataset:
- Kuvera-4B & 8B: These are my upgraded non-reasoning models, fine-tuned to provide practical financial advice. I've specifically trained the 8B model to better understand the user's emotional context.
- Kuvera-12B: A first experimental reasoning model focused on the query resolution.
As the sole person working on this project, this release is a noticeable step forward from my previous work, offering more powerful and nuanced tools for financial AI.
I am actively looking to collaborate with others who are passionate about analyzing and improving the quality of personal finance advice generated by large language models. If this sounds like you, please reach out!
You can check these out on the following links:
Models:
- Akhil-Theerthala/Kuvera-8B-qwen3-v0.2.1
- Akhil-Theerthala/Kuvera-4B-unsloth-gemma3
- Akhil-Theerthala/kuvera-12B-v0.2.0-unsloth-gemma3
Dataset:
- Akhil-Theerthala/Kuvera-PersonalFinance-V2.1
P.S. The paper on the framework used to generate these models along with the detailed evaluation of the main 8B model's responses is going to be released soon!
What's new:
I've expanded the Personal Finance Reasoning Dataset, which now includes 18.9k samples of real-world financial questions paired with detailed, empathetic answers. The previous generation pipeline was also streamlined with better psychological context and response validations.
I've also released new Kuvera models trained on this improved dataset:
- Kuvera-4B & 8B: These are my upgraded non-reasoning models, fine-tuned to provide practical financial advice. I've specifically trained the 8B model to better understand the user's emotional context.
- Kuvera-12B: A first experimental reasoning model focused on the query resolution.
As the sole person working on this project, this release is a noticeable step forward from my previous work, offering more powerful and nuanced tools for financial AI.
I am actively looking to collaborate with others who are passionate about analyzing and improving the quality of personal finance advice generated by large language models. If this sounds like you, please reach out!
You can check these out on the following links:
Models:
- Akhil-Theerthala/Kuvera-8B-qwen3-v0.2.1
- Akhil-Theerthala/Kuvera-4B-unsloth-gemma3
- Akhil-Theerthala/kuvera-12B-v0.2.0-unsloth-gemma3
Dataset:
- Akhil-Theerthala/Kuvera-PersonalFinance-V2.1
P.S. The paper on the framework used to generate these models along with the detailed evaluation of the main 8B model's responses is going to be released soon!
-
- 2 replies
That seriously would be very helpful!
reacted to
Tonic's
post with ๐
7 months ago
Post
3407
๐๐ปโโ๏ธ Normalize adding compute & runtime traces to your model cards
-
- 2 replies
reacted to
DualityAI-RebekahBogdanoff's
post with โค๏ธ
8 months ago
Post
3619
As part of Duality AIโs recent Kaggle competition, weโve released a free, fully customizable cloud scenario designed to help you create targeted datasets with YOLO-compatible labels.
The cloud simulation lets you customize the:
๐ธ camera distance
๐๏ธ film grain variation
๐ผ๏ธbackground objects,
โ and more!
Create the dataset that you need by following this link: https://falcon.duality.ai/secure/scenarios/edit/cca0bc47-265a-4f67-843f-a434b63271b3?utm_source=huggingface&utm_medium=social&utm_campaign=general
Iโve attached an instructional video we used for the competition, but this feature is free for anyone who has an account. https://vimeo.com/1091271731?share=copy
The cloud simulation lets you customize the:
๐ธ camera distance
๐๏ธ film grain variation
๐ผ๏ธbackground objects,
โ and more!
Create the dataset that you need by following this link: https://falcon.duality.ai/secure/scenarios/edit/cca0bc47-265a-4f67-843f-a434b63271b3?utm_source=huggingface&utm_medium=social&utm_campaign=general
Iโve attached an instructional video we used for the competition, but this feature is free for anyone who has an account. https://vimeo.com/1091271731?share=copy
posted
an
update
8 months ago
Post
1195
Kuvera v0.1.0 is now live!
A series of personal finance advisor models that try to resolve the queries by trying to understand the personโs psychological state and relevant context.
These are still prototypes that have much room for improvement.
Whatโs included in this release:
- Akhil-Theerthala/Kuvera-8B-v0.1.0: Qwen3-8B, meticulously fine-tuned on approximately 20,000 personal-finance inquiries.
- Akhil-Theerthala/Kuvera-14B-v0.1.0: LoRA on DeepSeek-R1-Distill-Qwen-14B, honed through training on about 10,000 chain-of-thought queries.
For those interested, the models and datasets are accessible for free (links in the comments). If you are curious about the upcoming version's roadmap, letโs connectโthere are many more developments I plan to make, and would definitely appreciate any help.
A series of personal finance advisor models that try to resolve the queries by trying to understand the personโs psychological state and relevant context.
These are still prototypes that have much room for improvement.
Whatโs included in this release:
- Akhil-Theerthala/Kuvera-8B-v0.1.0: Qwen3-8B, meticulously fine-tuned on approximately 20,000 personal-finance inquiries.
- Akhil-Theerthala/Kuvera-14B-v0.1.0: LoRA on DeepSeek-R1-Distill-Qwen-14B, honed through training on about 10,000 chain-of-thought queries.
For those interested, the models and datasets are accessible for free (links in the comments). If you are curious about the upcoming version's roadmap, letโs connectโthere are many more developments I plan to make, and would definitely appreciate any help.
reacted to
shukdevdatta123's
post with ๐
8 months ago
Post
3056
Excited to share my latest project: an AI-powered Educational Content Creator Assistant! ๐โจ Built with Gradio and OpenAI, it transforms PDF/DOCX documents into engaging study materials like interactive flashcards, quizzes, summaries, and lesson plans. Features GPU acceleration and downloadable HTML outputs. Perfect for educators and students! ๐ #EdTech #AI #Python #ML #LLM
Features Include:
- โ AI-powered content generation
- โ Document processing (PDF/DOCX)
- โ Interactive flashcards with slider functionality
- โ Comprehensive summaries
- โ Structured study notes
- โ Quiz generation with multiple choice, short answer, and essay questions
- โ Mind map structure creation
- โ Detailed lesson plans
- โ In-depth concept explanations
- โ Practice problems (beginner to challenge levels)
- โ Downloadable HTML outputs with interactivity
- โ Gradio-based user interface
- โ OpenAI API integration via OpenRouter
- โ GPU acceleration with ZeroGPU
- โ Real-time status updates for API and document processing
- โ Support for multiple content types
- โ Formatted HTML content with CSS styling
- โ Print functionality in downloadable files
- โ Error handling for document processing and API calls
- โ Modular function design for content generation
Youtube: https://www.youtube.com/watch?v=4a52HcioWPk
Demo: https://shukdevdatta123-ecca.hf.space
Features Include:
- โ AI-powered content generation
- โ Document processing (PDF/DOCX)
- โ Interactive flashcards with slider functionality
- โ Comprehensive summaries
- โ Structured study notes
- โ Quiz generation with multiple choice, short answer, and essay questions
- โ Mind map structure creation
- โ Detailed lesson plans
- โ In-depth concept explanations
- โ Practice problems (beginner to challenge levels)
- โ Downloadable HTML outputs with interactivity
- โ Gradio-based user interface
- โ OpenAI API integration via OpenRouter
- โ GPU acceleration with ZeroGPU
- โ Real-time status updates for API and document processing
- โ Support for multiple content types
- โ Formatted HTML content with CSS styling
- โ Print functionality in downloadable files
- โ Error handling for document processing and API calls
- โ Modular function design for content generation
Youtube: https://www.youtube.com/watch?v=4a52HcioWPk
Demo: https://shukdevdatta123-ecca.hf.space
reacted to
onekq's
post with โค๏ธโค๏ธ
8 months ago
Post
2213
Highly recommend the latest Gemini Flash. My favorite Google I/O gift. It ranks behind reasoning models but runs a lot faster than them. It beats DeepSeek v3.
onekq-ai/WebApp1K-models-leaderboard
Reasoning is good for coding, but not mandatory.
onekq-ai/WebApp1K-models-leaderboard
Reasoning is good for coding, but not mandatory.
- 1 reply
reacted to
KaraKaraWitch's
post with ๐ฅ
8 months ago
Post
2768
> New Model
> Looks at Model Card
> "Open-Weights"
> Looks at Model Card
> "Open-Weights"
- 1 reply
reacted to
merve's
post with ๐ฅ
9 months ago
Post
6725
A real-time object detector much faster and accurate than YOLO with Apache 2.0 license just landed to Hugging Face transformers ๐ฅ
D-FINE is the sota real-time object detector that runs on T4 (free Colab) ๐คฉ
> Collection with all checkpoints and demo ustc-community/d-fine-68109b427cbe6ee36b4e7352
Notebooks:
> Tracking https://github.com/qubvel/transformers-notebooks/blob/main/notebooks/DFine_tracking.ipynb
> Inference https://github.com/qubvel/transformers-notebooks/blob/main/notebooks/DFine_inference.ipynb
> Fine-tuning https://github.com/qubvel/transformers-notebooks/blob/main/notebooks/DFine_finetune_on_a_custom_dataset.ipynb
h/t @vladislavbro @qubvel-hf @ariG23498 and the authors of the paper ๐ฉ
Regular object detectors attempt to predict bounding boxes in (x, y, w, h) pixel perfect coordinates, which is very rigid and hard to solve ๐ฅฒโน๏ธ
D-FINE formulates object detection as a distribution for bounding box coordinates, refines them iteratively, and it's more accurate ๐คฉ
Another core idea behind this model is Global Optimal Localization Self-Distillation โคต๏ธ
this model uses final layer's distribution output (sort of like a teacher) to distill to earlier layers to make early layers more performant.
D-FINE is the sota real-time object detector that runs on T4 (free Colab) ๐คฉ
> Collection with all checkpoints and demo ustc-community/d-fine-68109b427cbe6ee36b4e7352
Notebooks:
> Tracking https://github.com/qubvel/transformers-notebooks/blob/main/notebooks/DFine_tracking.ipynb
> Inference https://github.com/qubvel/transformers-notebooks/blob/main/notebooks/DFine_inference.ipynb
> Fine-tuning https://github.com/qubvel/transformers-notebooks/blob/main/notebooks/DFine_finetune_on_a_custom_dataset.ipynb
h/t @vladislavbro @qubvel-hf @ariG23498 and the authors of the paper ๐ฉ
Regular object detectors attempt to predict bounding boxes in (x, y, w, h) pixel perfect coordinates, which is very rigid and hard to solve ๐ฅฒโน๏ธ
D-FINE formulates object detection as a distribution for bounding box coordinates, refines them iteratively, and it's more accurate ๐คฉ
Another core idea behind this model is Global Optimal Localization Self-Distillation โคต๏ธ
this model uses final layer's distribution output (sort of like a teacher) to distill to earlier layers to make early layers more performant.
- 2 replies