Deathsquad10/TinyLlama-1.1B-Remix-V.2

Model Overview

Deathsquad10/TinyLlama-1.1B-Remix-V.2 is a 1.1 billion parameter language model built on the Llama 2 architecture and tokenizer. It is a chat-tuned version, fine-tuned from the TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T base model.

Key Capabilities & Training

This model was developed following a training recipe similar to Hugging Face's Zephyr. Its fine-tuning process involved:

• Initial fine-tuning on a variant of the UltraChat dataset, which provides diverse synthetic dialogues.
• Further alignment using 🤗 TRL's DPOTrainer on the openbmb/UltraFeedback dataset, which contains 64k prompts and GPT-4 ranked model completions.

Performance

While compact, the model demonstrates general language understanding capabilities, as indicated by its zero-shot performance on various tasks:

• ARC-Challenge: 0.2790 acc, 0.3234 acc_norm
• ARC-Easy: 0.6006 acc, 0.5770 acc_norm
• BoolQ: 0.6373 acc
• HellaSwag: 0.4521 acc, 0.5822 acc_norm
• PIQA: 0.7269 acc, 0.7296 acc_norm

Use Cases

Given its compact size (1.1B parameters) and chat-specific fine-tuning, this model is well-suited for:

• Applications with restricted computational and memory footprints.
• Integration into projects built upon the Llama ecosystem due to architectural compatibility.
• Conversational AI tasks where a smaller, efficient chat model is preferred.