GPT2-Instruct
A fully on-chain GPT2-Instruct implementation – both front-end and backend hosted on the Internet Computer. This free version, v0.2, can process up to 56 tokens – input and output. Although the output it still deterministic, the temperature is now adjustable currently with 0.7 as the default value.
Disclaimer
This is an early GPT-2 implementation. Responses may be inaccurate or inconsistent. Use for demonstration purposes only. Your feedback helps us improve!
User:
Assistant:
Temperature:
0.7
Tokenization
Canister:
inb6i-gaaaa-aaaao-a3tna-cai
Query:
tokenize: (text) → (OK: (vec nat32, vec text), Err: text) query
Info
This canister handles the tokenization process for our GPT-2 implementation. It converts raw text inputs into tokens that can be processed by the model.
Inference
Canister:
ikay4-lyaaa-aaaao-a3tnq-cai
Query:
inference: (vec nat32, nat8, float32) -> (Ok: (Ok: vec nat32, Err: text), opt blob); Err: text) query
Info
The inference canister is responsible for running the actual GPT-2 model. It takes the tokenized input and generates the model’s output responses.
