🐛 Stop bigcode prompts (#170)

.env

@@ -13,6 +13,7 @@ MODELS=`[
     "websiteUrl": "https://open-assistant.io",
     "userMessageToken": "<|prompter|>",
     "assistantMessageToken": "<|assistant|>",
+    "messageEndToken": "</s>",
     "preprompt": "Below are a series of dialogues between various people and an AI assistant. The AI tries to be helpful, polite, honest, sophisticated, emotionally aware, and humble-but-knowledgeable. The assistant is happy to help with almost anything, and will do its best to understand exactly what is needed. It also tries to avoid giving false or misleading information, and it caveats when it isn't entirely sure about the right answer. That said, the assistant is practical and really does its best, and doesn't let caution get too much in the way of being useful.\n-----\n",
     "promptExamples": [
       {
@@ -33,7 +34,6 @@ MODELS=`[
       "top_k": 50,
       "truncate": 1000,
       "max_new_tokens": 1024,
-      "stop":["</s>"]
     }
   },
   {
@@ -61,7 +61,8 @@ MODELS=`[
       "top_p": 0.9,
       "repetition_penalty": 1.2,
       "truncate": 8000,
-      "max_new_tokens": 2000
+      "max_new_tokens": 2000,
+      "stop": ["Human:", "-----"]
     }
   }
 ]`

src/lib/buildPrompt.ts

@@ -17,10 +17,10 @@ export function buildPrompt(
 					(m.from === "user"
 						? model.userMessageToken + m.content
 						: model.assistantMessageToken + m.content) +
-					(model.parameters.stop
-						? m.content.endsWith(model.parameters.stop[0])
+					(model.messageEndToken
+						? m.content.endsWith(model.messageEndToken)
 							? ""
-							: model.parameters.stop[0]
+							: model.messageEndToken
 						: "")
 			)
 			.join("") + model.assistantMessageToken;

src/lib/components/chat/ChatMessage.svelte

@@ -8,22 +8,32 @@
 	import IconLoading from "../icons/IconLoading.svelte";
 	import CarbonRotate360 from "~icons/carbon/rotate-360";
 	import { PUBLIC_SEP_TOKEN } from "$lib/constants/publicSepToken";
+	import type { Model } from "$lib/types/Model";
 
 	function sanitizeMd(md: string) {
-		return md
+		let ret = md
 			.replace(/<\|[a-z]*$/, "")
 			.replace(/<\|[a-z]+\|$/, "")
 			.replace(/<$/, "")
 			.replaceAll(PUBLIC_SEP_TOKEN, " ")
 			.replaceAll(/<\|[a-z]+\|>/g, " ")
 			.replaceAll(/<br\s?\/?>/gi, "\n")
-			.trim()
-			.replaceAll("<", "&lt;");
+			.replaceAll("<", "&lt;")
+			.trim();
+
+		for (const stop of model.parameters.stop ?? []) {
+			if (ret.endsWith(stop)) {
+				ret = ret.slice(0, -stop.length).trim();
+			}
+		}
+
+		return ret;
 	}
 	function unsanitizeMd(md: string) {
 		return md.replaceAll("&lt;", "<");
 	}
 
+	export let model: Model;
 	export let message: Message;
 	export let loading = false;
 

src/lib/components/chat/ChatMessages.svelte

@@ -42,6 +42,7 @@
 			<ChatMessage
 				loading={loading && i === messages.length - 1}
 				{message}
+				model={currentModel}
 				on:retry={() => dispatch("retry", { id: message.id, content: message.content })}
 			/>
 		{:else}
@@ -50,7 +51,10 @@
 			{/if}
 		{/each}
 		{#if pending}
-			<ChatMessage message={{ from: "assistant", content: "", id: randomUUID() }} />
+			<ChatMessage
+				message={{ from: "assistant", content: "", id: randomUUID() }}
+				model={currentModel}
+			/>
 		{/if}
 		<div class="h-32 flex-none" />
 	</div>

src/lib/server/models.ts

@@ -10,6 +10,7 @@ const modelsRaw = z
 			datasetName: z.string().min(1).optional(),
 			userMessageToken: z.string().min(1),
 			assistantMessageToken: z.string().min(1),
+			messageEndToken: z.string().min(1).optional(),
 			preprompt: z.string().default(""),
 			prepromptUrl: z.string().url().optional(),
 			promptExamples: z
@@ -34,7 +35,7 @@ const modelsRaw = z
 					temperature: z.number().min(0).max(1),
 					truncate: z.number().int().positive(),
 					max_new_tokens: z.number().int().positive(),
-					stop: z.array(z.string()).min(1).optional(),
+					stop: z.array(z.string()).optional(),
 				})
 				.passthrough(),
 		})

src/lib/types/Model.ts

@@ -1,7 +1,6 @@
-export interface Model {
-	name: string;
-	displayName?: string;
-	websiteUrl?: string;
-	datasetName?: string;
-	promptExamples?: Array<{ title: string; prompt: string }>;
-}
+import type { BackendModel } from "$lib/server/models";
+
+export type Model = Pick<
+	BackendModel,
+	"name" | "displayName" | "websiteUrl" | "datasetName" | "promptExamples" | "parameters"
+>;

src/routes/+layout.server.ts

@@ -33,8 +33,7 @@ export const load: LayoutServerLoad = async ({ locals, depends, url }) => {
 		settings: {
 			shareConversationsWithModelAuthors: settings?.shareConversationsWithModelAuthors ?? true,
 			ethicsModalAcceptedAt: settings?.ethicsModalAcceptedAt ?? null,
-			activeModel:
-				url.searchParams.get("model") ?? settings?.activeModel ?? defaultModel.name,
+			activeModel: url.searchParams.get("model") ?? settings?.activeModel ?? defaultModel.name,
 		},
 		models: models.map((model) => ({
 			name: model.name,

src/routes/conversation/[id]/+server.ts

@@ -107,7 +107,16 @@ export async function POST({ request, fetch, locals, params }) {
 			generated_text = generated_text.slice(prompt.length);
 		}
 
-		generated_text = trimSuffix(trimPrefix(generated_text, "<|startoftext|>"), PUBLIC_SEP_TOKEN);
+		generated_text = trimSuffix(
+			trimPrefix(generated_text, "<|startoftext|>"),
+			PUBLIC_SEP_TOKEN
+		).trim();
+
+		for (const stop of modelInfo?.parameters?.stop ?? []) {
+			if (generated_text.endsWith(stop)) {
+				generated_text = generated_text.slice(0, -stop.length).trim();
+			}
+		}
 
 		messages.push({ from: "assistant", content: generated_text, id: crypto.randomUUID() });