Add seed, top-K, repeat penalty, and stop sequence generation parameters

Author: voidstackloop

README.md

@@ -33,7 +33,7 @@ Beyond chat, Modelforge includes an **agentic mode** — the model can read/writ
 
 **Organization**
 - **Projects** — group related chats under shared instructions and default model parameters.
-- **Per-session and per-project overrides** — pin a specific prompt, model, context length, or temperature to a single chat or an entire project, falling back to sane defaults.
+- **Per-session and per-project overrides** — pin a specific prompt, model, temperature, seed, top-K/top-P, repeat penalty, context length, GPU offload, or stop sequences to a single chat or an entire project, falling back to sane defaults. Provider-specific parameters (e.g. seed isn't supported by Claude, top-K isn't supported by ChatGPT) are automatically disabled when they don't apply to the selected model.
 - **Prompt library** — save and reuse system prompts across chats.
 - **Command palette** (`Ctrl/Cmd+K`) — jump between chats, projects, and settings without touching the mouse.
 

app/src/ollama-manager.ts

@@ -232,6 +232,10 @@ export async function chat(
                     presence_penalty: options?.presencePenalty,
                     num_ctx: options?.contextLength,
                     num_gpu: options?.gpuLayers,
+                    seed: options?.seed,
+                    top_k: options?.topK,
+                    repeat_penalty: options?.repeatPenalty,
+                    stop: options?.stop,
                 },
             }),
             signal,

app/src/providers/anthropic.ts

@@ -63,6 +63,8 @@ export const chat: ChatFn = async (apiKey, model, messages, options, onToken, si
                 stream: true,
                 temperature: options?.temperature ?? 0.7,
                 top_p: options?.topP,
+                top_k: options?.topK,
+                stop_sequences: options?.stop && options.stop.length > 0 ? options.stop : undefined,
                 ...(tools && tools.length > 0 ? { tools: toAnthropicTools(tools) } : {}),
             }),
             signal,

app/src/providers/openai.ts

@@ -57,6 +57,8 @@ export const chat: ChatFn = async (apiKey, model, messages, options, onToken, si
                 max_tokens: options?.maxTokens,
                 frequency_penalty: options?.frequencyPenalty,
                 presence_penalty: options?.presencePenalty,
+                seed: options?.seed,
+                stop: options?.stop && options.stop.length > 0 ? options.stop : undefined,
                 ...(tools && tools.length > 0 ? { tools: toOpenAiTools(tools) } : {}),
             }),
             signal,

app/src/providers/types.ts

@@ -61,6 +61,16 @@ export interface ChatOptions {
     // number = offload that many layers (useful for tuning multi-GPU setups
     // or freeing VRAM for something else running alongside Ollama).
     gpuLayers?: number;
+    // Ollama + OpenAI only (Anthropic has no reproducibility param). Same
+    // seed + same prompt should produce the same output, useful for testing.
+    seed?: number;
+    // Ollama + Anthropic only (OpenAI doesn't expose top-k sampling).
+    topK?: number;
+    // Ollama-only: penalizes tokens that already appeared recently, distinct
+    // from (and in addition to) the OpenAI-style frequency/presence penalties.
+    repeatPenalty?: number;
+    // All providers: stop generation as soon as any of these strings appears.
+    stop?: string[];
 }
 
 export type ChatFn = (

app/src/settings-store.ts

@@ -19,6 +19,10 @@ export interface AppSettings {
     contextLength: number;
     // undefined = auto (let Ollama decide how many layers to offload to GPU).
     gpuLayers?: number;
+    seed?: number;
+    topK?: number;
+    repeatPenalty?: number;
+    stop?: string[];
     systemPrompt: string;
     promptPresets: PromptPreset[];
     theme: "light" | "dark" | "system";

frontend/src/components/layout.tsx

@@ -313,6 +313,64 @@ function ProjectGroup({
                                         className="h-8 text-xs"
                                     />
                                 </div>
+                                <div className="flex flex-col gap-1">
+                                    <label className="text-xs text-muted-foreground">{t.seed}</label>
+                                    <Input
+                                        type="number"
+                                        step={1}
+                                        placeholder={t.seedRandom}
+                                        title={t.seedHelp}
+                                        value={params.seed ?? ""}
+                                        onChange={(e) => updateParam({ seed: e.target.value === "" ? undefined : Number(e.target.value) })}
+                                        aria-label={t.seed}
+                                        className="h-8 text-xs"
+                                    />
+                                </div>
+                                <div className="flex flex-col gap-1">
+                                    <label className="text-xs text-muted-foreground">{t.topK}</label>
+                                    <Input
+                                        type="number"
+                                        min={1}
+                                        step={1}
+                                        title={t.topKHelp}
+                                        value={params.topK ?? ""}
+                                        onChange={(e) => updateParam({ topK: e.target.value === "" ? undefined : Number(e.target.value) })}
+                                        aria-label={t.topK}
+                                        className="h-8 text-xs"
+                                    />
+                                </div>
+                                <div className="flex flex-col gap-1">
+                                    <label className="text-xs text-muted-foreground">{t.repeatPenalty}</label>
+                                    <Input
+                                        type="number"
+                                        min={0}
+                                        step={0.05}
+                                        title={t.repeatPenaltyHelp}
+                                        value={params.repeatPenalty ?? ""}
+                                        onChange={(e) =>
+                                            updateParam({ repeatPenalty: e.target.value === "" ? undefined : Number(e.target.value) })
+                                        }
+                                        aria-label={t.repeatPenalty}
+                                        className="h-8 text-xs"
+                                    />
+                                </div>
+                            </div>
+                            <div className="flex flex-col gap-1">
+                                <label className="text-xs text-muted-foreground">{t.stopSequences}</label>
+                                <Input
+                                    placeholder={t.stopSequencesPlaceholder}
+                                    value={(params.stop ?? []).join(", ")}
+                                    onChange={(e) =>
+                                        updateParam({
+                                            stop: e.target.value
+                                                .split(",")
+                                                .map((s) => s.trim())
+                                                .filter(Boolean),
+                                        })
+                                    }
+                                    aria-label={t.stopSequences}
+                                    className="h-8 text-xs"
+                                />
                             </div>
                             <p className="text-xs text-muted-foreground">
                                 Blank fields fall back to the global default. {t.penaltyClaudeNote}

frontend/src/lib/translations.ts

@@ -52,6 +52,16 @@ export interface Dictionary {
     gpuLayers: string;
     gpuLayersAuto: string;
     gpuLayersHelp: string;
+    seed: string;
+    seedRandom: string;
+    seedHelp: string;
+    topK: string;
+    topKHelp: string;
+    repeatPenalty: string;
+    repeatPenaltyHelp: string;
+    stopSequences: string;
+    stopSequencesHelp: string;
+    stopSequencesPlaceholder: string;
     contextLengthOllamaOnly: string;
     frequencyPenalty: string;
     presencePenalty: string;
@@ -142,6 +152,16 @@ export const en: Dictionary = {
     gpuLayers: "GPU layers",
     gpuLayersAuto: "Auto",
     gpuLayersHelp: "How many model layers to offload to the GPU (Ollama's num_gpu). Leave blank to let Ollama decide automatically; 0 forces CPU-only.",
+    seed: "Seed",
+    seedRandom: "Random",
+    seedHelp: "A fixed seed makes output reproducible — same seed and prompt should give the same response. Leave blank for a random seed each time. Not supported by Claude.",
+    topK: "Top K",
+    topKHelp: "Limits sampling to the K most likely next tokens. Lower values are more focused/deterministic. Not supported by ChatGPT.",
+    repeatPenalty: "Repeat penalty",
+    repeatPenaltyHelp: "Penalizes tokens that already appeared recently, reducing repetition. 1.0 = no penalty. Ollama only.",
+    stopSequences: "Stop sequences",
+    stopSequencesHelp: "Generation stops as soon as any of these strings appears in the output. Comma-separated.",
+    stopSequencesPlaceholder: "e.g. \\n\\nUser:, ###, <|end|>",
     contextLengthOllamaOnly: " (Ollama only)",
     frequencyPenalty: "Frequency penalty",
     presencePenalty: "Presence penalty",
@@ -234,6 +254,16 @@ export const tr: Dictionary = {
     gpuLayers: "GPU katmanları",
     gpuLayersAuto: "Otomatik",
     gpuLayersHelp: "GPU'ya kaç model katmanının aktarılacağı (Ollama'nın num_gpu ayarı). Ollama'nın otomatik karar vermesi için boş bırakın; 0 sadece CPU kullanımını zorlar.",
+    seed: "Seed",
+    seedRandom: "Rastgele",
+    seedHelp: "Sabit bir seed, çıktıyı tekrarlanabilir kılar — aynı seed ve istemle aynı yanıt alınır. Her seferinde rastgele bir seed için boş bırakın. Claude tarafından desteklenmez.",
+    topK: "Top K",
+    topKHelp: "Örneklemeyi en olası K sonraki token ile sınırlar. Düşük değerler daha odaklı/deterministik sonuç verir. ChatGPT tarafından desteklenmez.",
+    repeatPenalty: "Tekrar cezası",
+    repeatPenaltyHelp: "Yakın zamanda geçmiş olan token'ları cezalandırarak tekrarı azaltır. 1.0 = ceza yok. Sadece Ollama.",
+    stopSequences: "Durdurma dizileri",
+    stopSequencesHelp: "Çıktıda bu dizilerden herhangi biri göründüğünde üretim durur. Virgülle ayırın.",
+    stopSequencesPlaceholder: "örn. \\n\\nUser:, ###, <|end|>",
     contextLengthOllamaOnly: " (yalnızca Ollama)",
     frequencyPenalty: "Sıklık cezası",
     presencePenalty: "Varlık cezası",

frontend/src/pages/Chat.tsx

@@ -507,6 +507,10 @@ export default function Chat() {
             presencePenalty: params.presencePenalty ?? project?.params?.presencePenalty ?? settings?.presencePenalty,
             contextLength: params.contextLength ?? project?.params?.contextLength ?? settings?.contextLength,
             gpuLayers: params.gpuLayers ?? project?.params?.gpuLayers ?? settings?.gpuLayers,
+            seed: params.seed ?? project?.params?.seed ?? settings?.seed,
+            topK: params.topK ?? project?.params?.topK ?? settings?.topK,
+            repeatPenalty: params.repeatPenalty ?? project?.params?.repeatPenalty ?? settings?.repeatPenalty,
+            stop: params.stop ?? project?.params?.stop ?? settings?.stop,
         };
     }
 
@@ -1070,6 +1074,72 @@ export default function Chat() {
                                     className="h-8 text-xs"
                                 />
                             </div>
+                            <div className="flex flex-col gap-1">
+                                <label className="text-xs text-muted-foreground">
+                                    {t.seed}
+                                    {parsedModel?.provider === "anthropic" ? " (not supported by Claude)" : ""}
+                                </label>
+                                <Input
+                                    type="number"
+                                    step={1}
+                                    placeholder={t.seedRandom}
+                                    title={t.seedHelp}
+                                    value={params.seed ?? currentProject?.params?.seed ?? settings?.seed ?? ""}
+                                    onChange={(e) => updateParam({ seed: e.target.value === "" ? undefined : Number(e.target.value) })}
+                                    disabled={parsedModel?.provider === "anthropic"}
+                                    className="h-8 text-xs"
+                                />
+                            </div>
+                            <div className="flex flex-col gap-1">
+                                <label className="text-xs text-muted-foreground">
+                                    {t.topK}
+                                    {parsedModel?.provider === "openai" ? " (not supported by ChatGPT)" : ""}
+                                </label>
+                                <Input
+                                    type="number"
+                                    min={1}
+                                    step={1}
+                                    title={t.topKHelp}
+                                    value={params.topK ?? currentProject?.params?.topK ?? settings?.topK ?? ""}
+                                    onChange={(e) => updateParam({ topK: e.target.value === "" ? undefined : Number(e.target.value) })}
+                                    disabled={parsedModel?.provider === "openai"}
+                                    className="h-8 text-xs"
+                                />
+                            </div>
+                            <div className="flex flex-col gap-1">
+                                <label className="text-xs text-muted-foreground">
+                                    {t.repeatPenalty}
+                                    {parsedModel?.provider !== "ollama" ? " (Ollama only)" : ""}
+                                </label>
+                                <Input
+                                    type="number"
+                                    min={0}
+                                    step={0.05}
+                                    title={t.repeatPenaltyHelp}
+                                    value={params.repeatPenalty ?? currentProject?.params?.repeatPenalty ?? settings?.repeatPenalty ?? ""}
+                                    onChange={(e) =>
+                                        updateParam({ repeatPenalty: e.target.value === "" ? undefined : Number(e.target.value) })
+                                    }
+                                    disabled={parsedModel?.provider !== "ollama"}
+                                    className="h-8 text-xs"
+                                />
+                            </div>
+                        </div>
+                        <div className="mt-3 flex flex-col gap-1">
+                            <label className="text-xs text-muted-foreground">{t.stopSequences}</label>
+                            <Input
+                                placeholder={t.stopSequencesPlaceholder}
+                                value={(params.stop ?? currentProject?.params?.stop ?? settings?.stop ?? []).join(", ")}
+                                onChange={(e) =>
+                                    updateParam({
+                                        stop: e.target.value
+                                            .split(",")
+                                            .map((s) => s.trim())
+                                            .filter(Boolean),
+                                    })
+                                }
+                                className="h-8 text-xs"
+                            />
                         </div>
                     </PopoverContent>
                 </Popover>