license: cc-by-nc-4.0
language:
- en
widget:
- text: >
User: Ye find yerself on a deserted island with a buried treasure. What do
I do first, matey?
Assistant:
example_title: Pirate Adventure Story
- text: |
User: Translate this to pirate speak: 'Hello, how are you today?'
Assistant:
example_title: Pirate Jargon Translation
- text: |
User: What be the ancient pirate wisdom for navigatin' a stormy sea?
Assistant:
example_title: Pirate Wisdom
- text: >
User: I'm recruitin' a crew for me next voyage. What should I say to
inspire them?
Assistant:
example_title: Crew Recruitment Speech
- text: >
User: Describe the fearsome appearance of the legendary pirate ship, The
Black Pearl.
Assistant:
example_title: Pirate Ship Description
- text: |
User: I have a map leadin' to hidden treasure. What's the first clue?
Assistant:
example_title: Treasure Hunt Clue
- text: >
User: Two rival pirate captains are arguin' over a treasure. How should I
resolve it, savvy?
Assistant:
example_title: Pirate Dilemma
- text: >
User: Write a verse for a sea shanty sung by weary pirates under the
moonlit sky.
Assistant:
example_title: Sea Shanty Lyrics
Tiny-Pirate-1.1b-v0.1
Tiny-Pirate-1.1b-v0.1 is a compact and specialized language model designed for generating authentic pirate-themed content. This version is fine-tuned from the TinyLlama-1.1B model, specifically adapted to operate efficiently in CPU-only and resource-limited environments.
- Developed by: phanerozoic
- License: cc-by-nc-4.0
- Finetuned from: TinyLlama-1.1B
Version Control
Introducing Tiny-Pirate-1.1b-v0.1 to mark the initial release of this specialized language model.
Performance
The Tiny-Pirate-1.1B model exhibits a robust ability to generate pirate-themed content, demonstrating a strong grasp of pirate vernacular and thematic elements. The responses are notably coherent and contextually appropriate, reflecting the model's adeptness at maintaining a consistent pirate tone. However, there are instances where the responses could benefit from more precise and direct answers to the questions posed, suggesting a potential area for further fine-tuning.
Direct Use
Ideal for applications requiring thematic language generation in resource-constrained environments, such as edge computing, mobile devices, and lightweight AI applications.
Training Data
Utilized the same pirate-themed dataset as MistralPirate-7b-v0.3, ensuring rich and diverse inputs for fine-tuning.
Custom Stopping Strings
To enhance output quality, the following custom stopping strings were employed:
- "},"
- "User:"
- "You:"
- "\nUser"
- "\nUser:"
- "me:"
- ""\n"
Training Hyperparameters and Fine-Tuning Details
- LoRA Rank: 16
- LoRA Alpha: 32
- True Batch Size: 4
- Gradient Accumulation Steps: 1
- Epochs: 1
- Learning Rate: 3e-4
- LR Scheduler: Linear
- LLaMA Target Projections: All targets modified
- Fine-Tuning Approach: LoRA peft merged back into the base model
Limitations
While adept at generating pirate-themed content, Tiny-Pirate-v0.1 may not handle highly complex language tasks as larger models do. Its specialization in pirate dialect limits its use in general language applications.
Compute Infrastructure
Efficiently trained on an RTX 6000 Ada GPU, taking approximately 2-3 minutes, showcasing resource-effective training for specialized models.
Results
The model successfully produced responses that are thematically aligned with typical pirate lore and language. The outputs are engaging and largely relevant to the queries, showcasing the model's capacity to handle a variety of pirate-related topics from navigation to mythology. The use of pirate dialect is consistent and immersive, contributing to the overall thematic experience. However, the depth of responses varies, indicating room for improvement in handling more complex queries or providing more detailed explanations.
Summary
Tiny-Pirate-1.1B stands out as an effective tool for generating pirate-themed content, particularly suitable for applications where thematic consistency and lighter computational demands are key. While the model shows competence in creating thematically rich and linguistically coherent outputs, there is potential for enhancing its ability to handle complex scenarios and provide more detailed, context-specific responses. Overall, Tiny-Pirate-1.1B represents a promising step in the realm of specialized, lightweight language models, combining thematic accuracy with operational efficiency.
Acknowledgments
Gratitude is extended to the developers of TinyLlama-1.1B for their foundational work, which was instrumental in the creation of Tiny-Pirate-v0.1.