ToolkenGPT

Source code for ToolkenGPT: Augmenting Frozen Language Models with Massive Tools via Tool Embeddings

NeurIPS 2023 (oral) | Best Paper Award at SoCalNLP 2023

Preparation

• Our experiments are conducted with LLaMA-13B/33B, which takes at least 2/4 GPUs of 24GB memory each.
• Acquire the checkpoints of LLaMA from MetaAI and install all required packages. Please refer to LLaMA official repo.
• Download the data from here (all datasets uploaded)
• (For VirtualHome) Please download the data following the instructions here.

  A side note: the folder virtualhome is from its official repo, but we fixed some small bugs in the evolving graph.

GSM8K-XL

Train

CUDA_VISIBLE_DEVICES=0,1,2,3 python -m torch.distributed.run --nproc_per_node 4 --master_port 1200 train_llama.py --ckpt_dir $LLAMA_CKPTS/30B --tokenizer_path $LLAMA_CKPTS/tokenizer.model --input_file data/gsm8k-xl/train.json --lr 1e-3 --num_epochs 10

Inference

CUDA_VISIBLE_DEVICES=0,1,2,3 python -m torch.distributed.run --nproc_per_node 4 --master_port 1250 inference_llama.py --ckpt_dir $LLAMA_CKPTS/30B --tokenizer_path $LLAMA_CKPTS/tokenizer.model --mode func_embedding --dataset gsm8k-xl  --func_load_path checkpoints/gsm8k-xl/epoch_3.pth --logits_bias 3.0

FuncQA

Train

CUDA_VISIBLE_DEVICES=0,1,2,3 python -m torch.distributed.run --nproc_per_node 4 --master_port 1200 train_llama.py --ckpt_dir $PATH_TO_LLAMA/30B --tokenizer_path $PATH_TO_LLAMA/tokenizer.model --input_file data/funcqa/train.json --lr 1e-4 --num_epochs 10

Inference (1-hop)

CUDA_VISIBLE_DEVICES=0,1,2,3 python -m torch.distributed.run --nproc_per_node 4 --master_port 1250 inference_llama.py --ckpt_dir $LLAMA_CKPTS/30B --tokenizer_path $LLAMA_CKPTS/tokenizer.model --mode func_embedding --dataset funcqa_oh --func_load_path checkpoints/funcqa/epoch_7.pth --logits_bias 2.7

Inference (MultiHop)

CUDA_VISIBLE_DEVICES=0,1,2,3 python -m torch.distributed.run --nproc_per_node 4 --master_port 1250 inference_llama.py --ckpt_dir $LLAMA_CKPTS/30B --tokenizer_path $LLAMA_CKPTS/tokenizer.model --mode func_embedding --dataset funcqa_mh --func_load_path checkpoints/funcqa/epoch_7.pth --logits_bias 4.0

VirtualHome

Training

python -m torch.distributed.run --nproc_per_node 2 --master_port 3001 train_llama.py --ckpt_dir $LLAMA_CKPTS/13B --tokenizer_path $LLAMA_CKPTS/tokenizer.model --dataset vh --input_file data/vh/legal_train_v4_embedding.json --only_functoken True --num_epochs 10

Inference

CUDA_VISIBLE_DEVICES=3,5 python -m torch.distributed.run --nproc_per_node 2 inference_llama.py --ckpt_dir $LLAMA_CKPTS/13B --tokenizer_path $LLAMA_CKPTS/tokenizer.model --mode vh_embedding_inference --dataset vh --func_load_path checkpoints/vh/epoch_7.pth --logits_bias 10.0

Evaluation

See evaluation/eval_vh.ipynb

KAMEL

Train

• synthetic data

  CUDA_VISIBLE_DEVICES=2,3 python -m torch.distributed.run --nproc_per_node 2 --master_port 3002 train_llama.py --ckpt_dir $LLAMA_CKPTS/13B --tokenizer_path $LLAMA_CKPTS/tokenizer.model --dataset kamel --input_file data/kamel/train_clean.json --only_functoken False ---log_every 500 --num_epochs 10

• supervised data

  CUDA_VISIBLE_DEVICES=2,3 python -m torch.distributed.run --nproc_per_node 2 --master_port 3002 train_llama.py --ckpt_dir $LLAMA_CKPTS/13B --tokenizer_path $LLAMA_CKPTS/tokenizer.model --dataset kamel --input_file data/kamel/kamel_id_train.json --only_functoken False ---log_every 500 --num_epochs 10

Inference

CUDA_VISIBLE_DEVICES=2,3 python -m torch.distributed.run --nproc_per_node 2 inference_llama.py --ckpt_dir $LLAMA_CKPTS/13B --tokenizer_path $LLAMA_CKPTS/tokenizer.model --mode kamel_embedding_inference --dataset kamel_30 --func_load_path checkpoints/kamel/epoch_4.pth --logits_bias 10

Evaluation

See evaluation/eval_kamel.ipynb