rainmana/TinyLlama-1.1B-Chat-v1.0-heretic
The rainmana/TinyLlama-1.1B-Chat-v1.0-heretic is a 1.1 billion parameter Llama-architecture model, derived from TinyLlama/TinyLlama-1.1B-Chat-v1.0. This version is a 'decensored' variant, created using the Heretic v1.3.0 tool, and is notable for its reduced refusal rate compared to the original model. It maintains a 2048 token context length and is designed for chat applications where less restrictive content filtering is desired.
Loading preview...
rainmana/TinyLlama-1.1B-Chat-v1.0-heretic Overview
This model is a 'decensored' version of the original TinyLlama/TinyLlama-1.1B-Chat-v1.0, created by rainmana using the Heretic v1.3.0 tool. It is a compact 1.1 billion parameter Llama-architecture model, maintaining the same architecture and tokenizer as Llama 2, allowing for broad compatibility with existing open-source projects.
Key Differentiators
- Decensored Nature: The primary distinction is its 'decensored' characteristic, achieved through a process called "Abliteration" with specific parameters applied to the original model.
- Reduced Refusal Rate: Performance metrics indicate a significantly lower refusal rate, with 2 refusals out of 100 compared to 6 out of 100 for the original TinyLlama-1.1B-Chat-v1.0.
- Reproducible: The model's creation process is reproducible, with details provided in the
reproducedirectory.
Training and Fine-tuning
The base TinyLlama model was pretrained on 3 trillion tokens. This chat model variant was initially fine-tuned on a version of the UltraChat dataset and further aligned using 🤗 TRL's DPOTrainer on the UltraFeedback dataset, following HF's Zephyr training recipe.
Use Cases
This model is suitable for chat applications where a smaller footprint and a less restrictive content policy are preferred. Its compact size makes it ideal for environments with limited computational and memory resources.