fireact_codellama_34b_lora

Website: FireAct Agent

FireAct Llama-2/CodeLlama

FireAct Llama/CodeLlama is a collection of fine-tuned generative text models for performing ReAct with external search tools. Links to other models can be found in the Index section.

Foundation Model Details

Note: As the foundation models, Llama-2 and CodeLlama, are developed by Meta, please also read the guidance and license on their website, Llama-2 and CodeLlama, before using FireAct models.

Model Developers System 2 Research, Cambridge LTL, Monash University, Princeton PLI.

Variations FireAct models including Llama-2-7B full fine-tuned models, and Llama-2-[7B,13B], CodeLlama-[7B,13B,34B] LoRA fine-tuned models. All released models are fine-tuned on multi-task (HotpotQA/StrategyQA/MMLU) and multi-type (ReAct/CoT/Reflexion) data.

Input Models input text only.

Output Models generate text only.

Index

Full Fine-tuned Model

FireAct Llama-2:

• fireact_llama_2_7b

LoRA Fine-tuned Model

FireAct Llama-2:

• fireact_llama_2_7b_lora
• fireact_llama_2_13b_lora

FireAct CodeLlama:

• fireact_codellama_7b_lora
• fireact_codellama_13b_lora
• fireact_codellama_34b_lora

LoRA Training procedure

The following bitsandbytes quantization config was used during training:

• load_in_8bit: True
• load_in_4bit: False
• llm_int8_threshold: 6.0
• llm_int8_skip_modules: None
• llm_int8_enable_fp32_cpu_offload: False
• llm_int8_has_fp16_weight: False
• bnb_4bit_quant_type: fp4
• bnb_4bit_use_double_quant: False
• bnb_4bit_compute_dtype: float32

Framework versions

• PEFT 0.4.0