Models

The Writer LLM service enables you to customize and use the Writer LLMs outlined below.

Model	Parameters	Availability	License	Notes
Palmyra Small	128m	HuggingFace	Apache-2.0	Improving language understanding by generative pre-training – arxiv
Palmyra 3B	3B	HuggingFace	Apache-2.0	Improving language understanding by generative pre-training – arxiv
Palmyra Base	5B	HuggingFace	Apache-2.0	Improving language understanding by generative pre-training – arxiv
Camel 🐪	5B	HuggingFace	Apache-2.0	InstructGPT
Palmyra Large	20B	HuggingFace	Apache-2.0	Improving language understanding by generative pre-training – arxiv
PalmyraMed	20B	HuggingFace	Apache-2.0	Palmyra-Med: Instruction-Based Fine-Tuning of LLMs Enhancing Medical Domain Performance
InstructPalmyra	20B	HuggingFace	Apache-2.0	InstructGPT
InstructPalmyra	30B	API, Writer Platform	Enterprise License	Training language models to follow instructions with human feedback
Palmyra-R	30B	API, Writer Platform	Enterprise License	Autoregressive language model with Retrieval-Augmented Generation
Palmyra-E	30B	API, Writer Platform	Enterprise License	Autoregressive language model
Silk Road	--	--	Enterprise License	+85K Context Length
Palmyra-X	43B	API, Writer Platform, On-premises	Enterprise License	Becoming self-instruct: introducing early stopping criteria for minimal instruct tuning
Palmyra-X	43B	API	Beta	32K context window
PalmyraMed	40B	API, Writer Platform, On-premises	Enterprise License	Palmyra-Med: Instruction-Based Fine-Tuning of LLMs Enhancing Medical Domain Performance

These large language models have been pre-trained on a massive amount of Internet text. Pre-training involves taking a mathematical model with random mathematical parameters (weights) and adjusting those weights iteratively in response to discrepancies between the model's output and a comparison point indicating the expected output. The most common training method for large language models is next-word prediction over massive amounts of text.

Differences between models

Palmyra Small

Palmyra Small is the fastest of Writer’s LLMs and can perform important tasks such as text parsing, simple classification, address correction, and keyword recognition. Providing more context drives better performance.

Good at: Text parsing, simple classification, address correction, and keyword recognition

Palmyra Base

Palmyra Base is extremely powerful as well as incredibly fast. This model excels at many nuanced tasks such as sentiment classification and summarization. Palmyra Base is also effective as a general service chatbot, answering questions and performing Q&A.

Competent in: complex classification, text sentiment, and summarization

Camel 🐪

Camel-5b is a trained large language model that follows instructions. Based on Palmyra-Base is trained on ~70k instruction & response fine tuning records generated by Writer Team from the InstructGPT paper, including brainstorming, classification, closed quality assurance, generation, information extraction, open quality assurance, and summarization.

Palmyra Large

Palmyra Large is the most capable model family, capable of performing any task that the other models can, often with less instruction. Palmyra Large is good at comprehending the text's intent, solving logic problems, and explaining character motivations.

Good at: Few-shots, cause and effect, and audience summarization

InstructPalmyra

InstructPalmyra is the most capable model. It can perform any tasks that the other models are able to, often with higher quality, longer output, and better instruction-following.

Good at: Zero-shots, cause and effect

Palmyra-R

Palmyra-R models are a general-purpose fine-tuning recipe for retrieval-augmented generation, combining pre-trained parametric and non-parametric memory for language generation.

Palmyra-E

It is more capable than the GPT-3 and GPT-3.5 models, able to perform more complex tasks, and comes in three flavors: General, Healthcare, and Fintech. It is available on-premise or via an API.

Silk Road

Same capabilities as the Palmyra-E mode but with ~80K context length. still in early testing stage