pull down to refresh

But, in my experience, LLMs do not always comply. Is that your language's responsibility to solve or mine?
My take:
[HUMAN LANG] --> [Lexer + MINI LLM] --> JSON + Prompt -> [BIG LLM] --> OUTPUT + Prompt --> [Verification LLM]
That is I think you need 3 stages: (a) Lexer + Mini LLM that transpiles HUMAN lang into structured JSON and also outputs a static prompt describing how to interpret this JSON. Both of which are then fed (b) into big LLM, (c) Output code (plus instruction prompt) is then fed back to another smallish LLM to verify that OUTPUT matches the "MUST" "SHOULD" "NEVER" directives.
[Mini LLM] = Something like gemma3-270m finetuned as a lexer / json compiler [BIG LLM] = Claude 4 or whatever [Verification LLM] = Could be something like Qwen3.
Json takes too many tokens, and will not compile to json. however, i have thought about it.
reply
0 sats \ 0 replies \ @freetx 6h
{ "react_dev": { "model": "gpt-4", "temperature": 0.2 }, "todo_rules": { "NEVER": [ "use_class_components", "mutate_state_directly" ], "MUST": [ "use_hooks", "include_add_todo", "include_delete_todo", "include_toggle_complete" ], "SHOULD": [ "use_local_storage", "handle_empty_input" ], "AVOID": [ "inline_functions", "magic_numbers" ] }, "generates working component": { "INPUT": "Create a React todo list", "EXPECT": "contains useState and retu" } }
reply