Introducing the Palmyra-mini family: Powerful, lightweight, and ready to reason!
•
58
AI & ML interests
AGI, LLMs, Knowledge Graph, Palmyra, Domain Specific LLM
Recent Activity
View all activity
Articles
kiranr
updated
a
model
2 days ago
Post
231
Introducing Palmyra-mini: Compact AI Models for Efficient Inference
The Palmyra-mini family from Writer includes three lightweight models designed for high performance and efficient inference. These models are ideal for developers looking to integrate AI capabilities without excessive computational overhead.
Model Variants
* palmyra-mini: A base model for general-purpose generative tasks, achieving 52.6% on Big Bench Hard (exact match).
* palmyra-mini-thinking-a: Optimized for complex logical reasoning with a Chain of Thought (CoT) approach, scoring 82.87% on GSM8K (strict match).
* palmyra-mini-thinking-b: Specialized for mathematical reasoning, achieving 92.5% on AMC23.
Technical Details
* All models are based on the Qwen architecture, compatible with popular inference frameworks like vLLM, SGLang, and TGI.
* "Thinking" models utilize CoT training for enhanced reasoning capabilities.
* GGUF and MLX quantizations are available for optimized performance.
For more information, including benchmark methodologies and detailed performance metrics, refer to our blog post: (https://huggingface.co/blog/Writer/announcing-palmyra-mini).
Model repos can be found here:
* Writer/palmyra-mini
* Writer/palmyra-mini-thinking-a
* Writer/palmyra-mini-thinking-b
Also check out a mobile implementation of palmyra-mini on iOS here to see a to see a working example of how inference can be incorporated on-device.(https://github.com/tsperes/palmyra-mini-mobile/)
The Palmyra-mini family from Writer includes three lightweight models designed for high performance and efficient inference. These models are ideal for developers looking to integrate AI capabilities without excessive computational overhead.
Model Variants
* palmyra-mini: A base model for general-purpose generative tasks, achieving 52.6% on Big Bench Hard (exact match).
* palmyra-mini-thinking-a: Optimized for complex logical reasoning with a Chain of Thought (CoT) approach, scoring 82.87% on GSM8K (strict match).
* palmyra-mini-thinking-b: Specialized for mathematical reasoning, achieving 92.5% on AMC23.
Technical Details
* All models are based on the Qwen architecture, compatible with popular inference frameworks like vLLM, SGLang, and TGI.
* "Thinking" models utilize CoT training for enhanced reasoning capabilities.
* GGUF and MLX quantizations are available for optimized performance.
For more information, including benchmark methodologies and detailed performance metrics, refer to our blog post: (https://huggingface.co/blog/Writer/announcing-palmyra-mini).
Model repos can be found here:
* Writer/palmyra-mini
* Writer/palmyra-mini-thinking-a
* Writer/palmyra-mini-thinking-b
Also check out a mobile implementation of palmyra-mini on iOS here to see a to see a working example of how inference can be incorporated on-device.(https://github.com/tsperes/palmyra-mini-mobile/)
dmytro-writer
authored
a
paper
7 months ago
shelly-writer
authored
a
paper
7 months ago
kiranr
authored
a
paper
7 months ago
wassemgtk
authored
a
paper
7 months ago
melisa
authored
a
paper
7 months ago
Post
3258
I’ve been diving into the iRoPE architecture from Llama 4—a game-changer for long-context models! It interleaves local attention (with RoPE) for short contexts and global attention (with inference-time temp scaling) for long-range reasoning, aiming for infinite context. I’m going to try writing iRoPE—who wants to help?
Code: https://github.com/wassemgtk/iRoPE-try/blob/main/iRoPE.ipynb
Code: https://github.com/wassemgtk/iRoPE-try/blob/main/iRoPE.ipynb
-
- 1 reply
Post
2137
For fun, a new project: SuperTokenizer! A BPE tokenizer trained on C4 to beat GPT-4. Byte-level, A100-powered, and open-source. Messing around with tokens!
https://github.com/wassemgtk/SuperTokenizer
https://github.com/wassemgtk/SuperTokenizer
- 1 reply
melisa
authored
a
paper
10 months ago
Post
1921
# GESAL: Real-Time Adaptation for LLMs
We’re excited to unveil **Graph-Enhanced Singular Adaptive Learning (GESAL)**, a framework that lets LLMs like
🔗 **Code**: [https://github.com/writer/AI-Adaptive-Learning-GESAL](https://github.com/writer/AI-Adaptive-Learning-GESAL)
---
## Why GESAL?
Static LLMs struggle to adapt without heavy retraining. GESAL solves this with:
- **SVF**: Adapts weights via \( W' = U (\Sigma \cdot z) V^T \), using few parameters.
- **Graph Memory**: Stores adaptations in nodes for scalability.
- **RL**: Updates via \( J(z) = \mathbb{E}[\log \pi_z(y|x) r] \) based on feedback.
---
## How It Works
Ask "How many R’s in ‘strawberry’?" If it says "2" and you say "no," GESAL learns to say "3" next time, avoiding repeats.
---
## Try It
Built with Hugging Face’s
Needs a Hugging Face token for Llama-3.2-1B.
---
## Results
GESAL hits 95% accuracy after 5 feedbacks vs. LoRA’s 70%. It’s efficient (~0.5M params) and scalable.
We’re excited to unveil **Graph-Enhanced Singular Adaptive Learning (GESAL)**, a framework that lets LLMs like
meta-llama/Llama-3.2-1B adapt in real time using user feedback. Check out the code and white paper on GitHub!🔗 **Code**: [https://github.com/writer/AI-Adaptive-Learning-GESAL](https://github.com/writer/AI-Adaptive-Learning-GESAL)
---
## Why GESAL?
Static LLMs struggle to adapt without heavy retraining. GESAL solves this with:
- **SVF**: Adapts weights via \( W' = U (\Sigma \cdot z) V^T \), using few parameters.
- **Graph Memory**: Stores adaptations in nodes for scalability.
- **RL**: Updates via \( J(z) = \mathbb{E}[\log \pi_z(y|x) r] \) based on feedback.
---
## How It Works
Ask "How many R’s in ‘strawberry’?" If it says "2" and you say "no," GESAL learns to say "3" next time, avoiding repeats.
---
## Try It
Built with Hugging Face’s
transformers:pip install transformers torch numpy
python Adaptive_Learning_(GESAL).pyNeeds a Hugging Face token for Llama-3.2-1B.
---
## Results
GESAL hits 95% accuracy after 5 feedbacks vs. LoRA’s 70%. It’s efficient (~0.5M params) and scalable.
-
- 15 replies
wassemgtk
authored
a
paper
11 months ago
dmytro-writer
authored
2
papers
11 months ago
kiranr
authored
a
paper
11 months ago
Post
1582
🔥 RAG in just a few lines of code?!
Try out our Hacker News Listener with new built-in RAG capabilities and Palmyra X 004 from the team at Writer!
This Writer Framework app:
- Scrapes up to 500 HN stories and comments
- Uploads them to a Knowledge Graph
- Enables interactive chat with the content using graph-based RAG
- Provides source attribution with every response
The best part? Setting up RAG is now incredibly simple - just a few lines of code to connect your Knowledge Graph as a tool with Palmyra X 004.
🤗 Space: samjulien/hacker-news-listener
💻 Code: https://github.com/writer/framework-tutorials/tree/main/hacker-news-social-listener
Try out our Hacker News Listener with new built-in RAG capabilities and Palmyra X 004 from the team at Writer!
This Writer Framework app:
- Scrapes up to 500 HN stories and comments
- Uploads them to a Knowledge Graph
- Enables interactive chat with the content using graph-based RAG
- Provides source attribution with every response
The best part? Setting up RAG is now incredibly simple - just a few lines of code to connect your Knowledge Graph as a tool with Palmyra X 004.
🤗 Space: samjulien/hacker-news-listener
💻 Code: https://github.com/writer/framework-tutorials/tree/main/hacker-news-social-listener
Post
3272
🔥 Introducing "Writing in the Margins (WiM)" - better inference pattern for long context LLMs that solves the Lost-in-the-Middle problem 🔥
Paper page: Writing in the Margins: Better Inference Pattern for Long Context Retrieval (2408.14906)
TL;DR
Make your model write "margin notes" as you chunk prefill the KV cache. Then ask it reread all notes before it speaks up.
Works with humans, works with AI 🤖
WiM leverages the chunked prefill of the key-value cache, which concurrently generates query-based extractive summaries at each step of the prefill that are subsequently reintegrated at the end of the computation. We term these intermediate outputs “margins”, drawing inspiration from the practice of making margin notes for improved comprehension of long contexts in human reading. We show that this technique, which adds only minimal additional computation, significantly improves LLMs long context reasoning capabilities.
Think: Every chunk has a chance to be attended to/ be at the end of the context at least once. 🎉
📊 Results:
- An average accuracy boost of 7.5% in multi-hop reasoning tasks like HotpotQA and MultiHop-RAG.
- Even a 30% increase in F1-score for summarisation-like tasks (CWE).
Plus, WiM fits seamlessly into interactive applications (think: progress bar!). It can provide real-time progress updates during data retrieval and integration, making it user-friendly and transparent - a stark contrast to feeding 1mln tokens to an LLMs and waiting 6 min for the first token. 🤯
👩💻🧑💻 Check it out and contribute to our open-source project here: https://github.com/writer/writing-in-the-margins
🧠 More about chunked prefill: https://docs.vllm.ai/en/latest/models/performance.html#chunked-prefill
Paper page: Writing in the Margins: Better Inference Pattern for Long Context Retrieval (2408.14906)
TL;DR
Make your model write "margin notes" as you chunk prefill the KV cache. Then ask it reread all notes before it speaks up.
Works with humans, works with AI 🤖
WiM leverages the chunked prefill of the key-value cache, which concurrently generates query-based extractive summaries at each step of the prefill that are subsequently reintegrated at the end of the computation. We term these intermediate outputs “margins”, drawing inspiration from the practice of making margin notes for improved comprehension of long contexts in human reading. We show that this technique, which adds only minimal additional computation, significantly improves LLMs long context reasoning capabilities.
Think: Every chunk has a chance to be attended to/ be at the end of the context at least once. 🎉
📊 Results:
- An average accuracy boost of 7.5% in multi-hop reasoning tasks like HotpotQA and MultiHop-RAG.
- Even a 30% increase in F1-score for summarisation-like tasks (CWE).
Plus, WiM fits seamlessly into interactive applications (think: progress bar!). It can provide real-time progress updates during data retrieval and integration, making it user-friendly and transparent - a stark contrast to feeding 1mln tokens to an LLMs and waiting 6 min for the first token. 🤯
👩💻🧑💻 Check it out and contribute to our open-source project here: https://github.com/writer/writing-in-the-margins
🧠 More about chunked prefill: https://docs.vllm.ai/en/latest/models/performance.html#chunked-prefill
- 2 replies
axel-writer
authored
a
paper
over 1 year ago
melisa
authored
a
paper
over 1 year ago
wassemgtk
authored
a
paper
over 1 year ago