phanerozoic's picture
Update README.md
b8c8809
|
raw
history blame
4.01 kB
metadata
license: cc-by-nc-4.0
language:
  - en

MistralPirate-7b-v0.3

Model Card

Description

MistralPirate-7b-v0.3 is a sophisticated language model fine-tuned for generating intricate and authentic pirate-themed content. This version, correcting our version control from v2 to v0.3, builds upon MistralPirate-7b-v2 and leverages advancements from Mistral Instruct v0.2. It shows improved performance in pirate dialect accuracy and perplexity scores.

  • Developed by: phanerozoic
  • License: cc-by-nc-4.0
  • Finetuned from: Mistral Instruct v0.2

Version Control Correction

Correcting version control to v0.3 to reflect developmental progression and enhancements over the previous version.

Comparative Analysis with Previous Model

MistralPirate-7b-v0.3 demonstrates notable improvements over its predecessor in several key areas:

  • Pirate Dialect: The new model uses richer and more immersive pirate vernacular, enhancing the thematic experience.
  • Technical Accuracy: It shows a deeper understanding of complex sailing scenarios, providing detailed and practical advice in response to intricate questions.
  • Language Coherence: The model maintains a consistent tone and style, effectively blending pirate jargon with technical expertise.

Direct Use

Ideal for interactive storytelling, gaming, advanced educational content, and conversational AI in pirate-themed settings.

Downstream Use

Suitable for tasks requiring detailed language generation and domain-specific knowledge, like advanced thematic content creation or immersive language learning.

Out-of-Scope Use

Not intended for general-purpose language modeling or non-pirate-themed contexts. Usage outside its specialization may result in suboptimal performance.

Bias, Risks, and Limitations

Limited by its training data, may inherit biases. Best used where pirate-themed language is appropriate, not for serious or sensitive communication.

Recommendations

Recommended for thematic contexts, with an understanding of its specialized focus. Not for accurate information outside pirate dialect specialization.

Custom Stopping Strings Usage

Custom stopping strings employed for output quality:

  • "},"
  • "User:"
  • "You:"
  • "\nUser"
  • "\nUser:"

Training Data

Trained on a vast dataset in ChatML format, ensuring diverse and rich inputs.

Preprocessing

Advanced preprocessing into ChatML format.

Training Hyperparameters and Fine-Tuning Details

  • Training Regime: FP32
  • Warmup Steps: 1
  • Per Device Train Batch Size: 2
  • Gradient Accumulation Steps: 64
  • Max Steps: 1500
  • Learning Rate: 0.00015
  • Logging Steps: 1
  • Save Steps: 1
  • Lora Alpha: 32
  • Dimension Count: 16
  • Specific Lora Fine-Tuning Run:
    • Step: 26
    • Loss: 1.4906
    • Learning Rate: 0.00019814951887490748
    • Epoch: 0.01

Speeds, Sizes, Times

Approximately 12 minutes training time on RTX 6000 Ada GPU.

Testing Data

Achieved a perplexity score of 5.17 against the Wikitext database.

Factors

Focus on language coherence, pirate dialect adherence, and technical accuracy.

Metrics

Primary metric: Perplexity. Qualitative assessments of dialect authenticity and technical content.

Results

Marked improvement in sophisticated output with authentic pirate tone. Lower perplexity score demonstrates enhanced language modeling.

Summary

Represents a significant advancement in domain-specific language modeling, excelling in complex, authentic pirate-themed content.

Model Architecture and Objective

Based on Mistral Instruct v0.2, fine-tuned for high coherence and technical accuracy in pirate-themed content.

Compute Infrastructure

Trained on RTX 6000 Ada GPU for efficient training and improved perplexity scores.

Hardware

  • Type: RTX 6000 Ada
  • Utilization: Approx. 12 minutes for training.

Acknowledgments

Gratitude to Mistral and Mistral Instruct v0.2 teams. Appreciation to the language modeling community for support in domain-specific model enhancement.