Kurofu/UHI_E3LLM

UHI-E3LLM Adapter Collection

This repository contains the PEFT/LoRA adapters used by UHI-E3LLM for monthly Urban Heat Island (UHI) intensity forecasting and factor-level explanation generation. The collection preserves all trainable stages needed to reconstruct the final model pipeline without redistributing the upstream 8B base weights.

The implementation, processed-data instructions, and evaluation workflows are available in the UHI-E3LLM code repository at https://github.com/ZJU-DAILY/UHI_E3LLM.

The processed dataset is archived at https://doi.org/10.5281/zenodo.22161925.

Artifact layout

Artifact	Path	Required base	Role
SFT epoch 0	sft/checkpoint-epoch0/	DeepSeek-R1-Distill-Llama-8B	End of SFT epoch 1
SFT epoch 1	sft/checkpoint-epoch1/	DeepSeek-R1-Distill-Llama-8B	End of SFT epoch 2
SFT epoch 2	sft/checkpoint-epoch2/	DeepSeek-R1-Distill-Llama-8B	End of SFT epoch 3
SFT epoch 3	sft/checkpoint-epoch3/	DeepSeek-R1-Distill-Llama-8B	End of SFT epoch 4
SFT epoch 4	sft/checkpoint-epoch4/	DeepSeek-R1-Distill-Llama-8B	End of SFT epoch 5
Fused SFT	SFT_fusion/	DeepSeek-R1-Distill-Llama-8B	Greedy-selected weighted adapter fusion
PPO epoch 0	ppo/epoch0/	Base model with the fused SFT adapter merged	End of PPO outer epoch 0
PPO epoch 1	ppo/epoch1/	Base model with the fused SFT adapter merged	End of PPO outer epoch 1
Distillation epoch 0	distillation/epoch0/	Fused model with PPO epoch 0 merged	Explanation adapter for policy epoch 0
Distillation epoch 1	distillation/epoch1/	Fused model with PPO epoch 1 merged	Explanation adapter for policy epoch 1

Evaluation usage

Clone the UHI-E3LLM code repository, place this adapter repository at Deepseek8b_Adapters in the code-repository root, install the locked environment, and run:

cd UHI_E3LLM
huggingface-cli download deepseek-ai/DeepSeek-R1-Distill-Llama-8B \
  --local-dir model/deepseek-8b
conda activate uhi-e3llm
GPU_ID=0 bash Deepseek8b_Adapters/run_all_stage_inference.sh