tokyotech-llm/Llama-3.1-8B-code-ablation-exp3-LR2.5e-5-MINLR2.5E-6-WD0.1-iter0007500

TEXT GENERATIONPricing:Input $0.2 / Cached $0.028 / Output $0.32Concurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Oct 19, 2024License:llama3.3Architecture:Transformer Featherless Exclusive Cold

The tokyotech-llm/Llama-3.1-8B-code-ablation-exp3-LR2.5e-5-MINLR2.5E-6-WD0.1-iter0007500 is an 8 billion parameter Llama-3.1 model continually pre-trained on 50 billion tokens, including 16% pylint-filtered Python code from The-Stack-v2 and 84% multilingual text. Developed by tokyotech-llm, this model focuses on evaluating the impact of pylint filtering on code generation performance. It is intended for research in code generation and text completion in English and Japanese, with a 32768 token context length.

Loading preview...

Model Overview

This model, developed by tokyotech-llm, is a Llama-3.1-8B variant that underwent continual pre-training on 50 billion tokens. Its primary purpose is to evaluate the performance impact of pylint-filtered Python code within the SwallowCode ablation experiments. The training dataset comprised 16% pylint-filtered Python code from The-Stack-v2 (Experiment 3) and 84% multilingual text, including significant portions of Japanese and English corpora.

Key Capabilities

  • Code Generation: Optimized for code generation tasks, particularly in Python, due to its specialized training data.
  • Multilingual Text Completion: Capable of text completion in both English and Japanese.
  • Research Focus: Designed for research into the effects of data filtering on large language model performance.

Training Details

The model was trained using Megatron-LM on 64 NVIDIA H100 GPUs. It utilizes a Llama-3 tokenizer and supports a sequence length of 8,192 tokens. Evaluation was conducted using lm-evaluation-harness and BigCodeBench, covering benchmarks like HumanEval, HumanEval+, MMLU, and GSM8K.

Intended Use

This model is best suited for:

  • Research purposes: Specifically for studying code generation and the impact of data curation techniques.
  • Text completion: For generating text in English and Japanese.
  • Code-related tasks: Leveraging its training on a significant volume of filtered Python code. It is not instruction-tuned.