Initial commit: Ollama Personal Trainer & Unsloth MoE Orchestrator Studio

studio-ref/src/types.ts

@@ -0,0 +1,222 @@
+export type ModelArch = "llama3" | "qwen2.5" | "mistral" | "gemma2" | "deepseek" | "phi4" | "smollm" | "custom";
+
+export type FineTuneMethod =
+  | "qlora"
+  | "lora_plus"
+  | "dora"
+  | "orpo"
+  | "dpo"
+  | "simpo"
+  | "kto"
+  | "galore"
+  | "neftune"
+  | "longlora";
+
+export type PruningMethod =
+  | "structured_layer"
+  | "head_pruning"
+  | "vocab_trim"
+  | "laser_svd"
+  | "wanda"
+  | "magnitude_dropout";
+
+export type MoEMethod =
+  | "moefication"
+  | "dare_ties"
+  | "slerp"
+  | "passthrough_franken"
+  | "task_arithmetic"
+  | "linear_average";
+
+export type GGUFQuantType =
+  | "Q4_K_M"
+  | "Q4_K_S"
+  | "Q5_K_M"
+  | "Q5_K_S"
+  | "Q8_0"
+  | "IQ4_XS"
+  | "IQ3_XXS"
+  | "IQ2_XS"
+  | "BF16"
+  | "FP16";
+
+export interface BaseModelInfo {
+  id: string;
+  name: string;
+  huggingFaceId: string;
+  ollamaName: string;
+  parametersBillion: number;
+  layers: number;
+  hiddenDim: number;
+  heads: number;
+  kvHeads: number;
+  vocabSize: number;
+  defaultContext: number;
+  architecture: ModelArch;
+  baseSizeGb: number;
+  q4SizeGb: number;
+  description: string;
+  recommendedFor4080Super: boolean;
+}
+
+export interface TrainingHyperparameters {
+  // LoRA / PEFT
+  lora_r: number;
+  lora_alpha: number;
+  lora_dropout?: number;
+  target_modules?: string[];
+  bias?: "none" | "all" | "lora_only";
+  use_dora?: boolean;
+  use_rslora?: boolean;
+  
+  // Optimizer & Scheduler
+  batch_size: number;
+  gradient_accumulation_steps: number;
+  learning_rate: number;
+  lr_scheduler?: "cosine" | "linear" | "constant" | "cosine_with_restarts";
+  warmup_ratio?: number;
+  warmup_steps?: number;
+  weight_decay?: number;
+  max_grad_norm?: number;
+  optimizer?: "adamw_8bit" | "paged_adamw_8bit" | "adamw_torch" | "galore_adamw";
+  
+  // Training Duration & Precision
+  epochs: number;
+  max_steps?: number;
+  max_seq_length: number;
+  precision?: "bfloat16" | "float16";
+  use_gradient_checkpointing?: boolean;
+  use_unsloth_fast_backprop?: boolean;
+  neftune_noise_alpha?: number;
+  
+  // Preference Alignment (for ORPO/DPO/SimPO)
+  preference_beta?: number;
+  simpo_gamma?: number;
+}
+
+export interface PruningConfig {
+  enabled: boolean;
+  pruneMethod?: PruningMethod;
+  methods?: PruningMethod[];
+  layerPruningRange: [number, number]; // e.g. prune layers 16 to 24
+  targetLayersCount?: number;
+  headsPrunePercentage?: number;
+  headPruningRatio?: number; // 0.0 - 0.5
+  vocabTrimTarget?: number;
+  vocabTargetTokens?: number; // e.g. 32000 from 128000
+  laserReductionRank?: number; // e.g. 32
+  repairLoRASteps?: number;
+  healingLoraSteps?: number;
+}
+
+export interface MoEConfig {
+  enabled: boolean;
+  method: MoEMethod;
+  numExperts: number;
+  topK: number;
+  routerType: "softmax" | "sinkhorn" | "switch";
+  expertSources: {
+    name: string;
+    modelId: string;
+    weight: number;
+    specialization: string;
+  }[];
+}
+
+export interface MCPToolDeclaration {
+  id: string;
+  name: string;
+  description: string;
+  serverName: string;
+  parametersSchema: {
+    type: "object";
+    properties: Record<string, { type: string; description: string; enum?: string[] }>;
+    required?: string[];
+  };
+  sampleCallsCount?: number;
+}
+
+export interface TrainingDataSample {
+  id: string;
+  instruction: string;
+  input?: string;
+  output: string;
+  system?: string;
+  category?: string;
+  difficulty?: string;
+  toolCalls?: {
+    name: string;
+    arguments: Record<string, any>;
+  }[];
+  simulatedToolResult?: string;
+  isMcpSample?: boolean;
+}
+
+export interface GGUFConfig {
+  quantization: GGUFQuantType;
+  contextLength: number;
+  templateFormat?: "llama3" | "chatml" | "mistral" | "alpaca" | "deepseek" | "gemma";
+  systemPrompt: string;
+  stopTokens?: string[];
+  temperature: number;
+  top_p?: number;
+  top_k?: number;
+  repeat_penalty?: number;
+  num_gpu_layers: number; // 999 for full 4080 Super offload
+  threads?: number;
+}
+
+export interface HardwarePreset {
+  name: string;
+  vramGb: number;
+  cudaCores: number;
+  tensorCores: number;
+  recommendedBatch: number;
+  recommendedSeqLen: number;
+  recommendedQuant: GGUFQuantType;
+  notes: string;
+}
+
+export interface DistillationConfig {
+  enabled: boolean;
+  teacherModel: string;
+  studentModel?: string;
+  distillationType?: "response_generation" | "cot_reasoning" | "logit_kl" | "mcp_alignment";
+  temperature: number;
+  includeThoughtChain: boolean;
+  distillDatasetSize?: number;
+  samplesToGenerate?: number;
+  distillationAlpha?: number;
+}
+
+export interface TrainingLogEntry {
+  step: number;
+  epoch: number;
+  loss: number;
+  evalLoss?: number;
+  learningRate: number;
+  gradNorm: number;
+  vramUsedGb: number;
+  tokensPerSec: number;
+  sampleOutput?: string;
+  timestamp: string;
+}
+
+export type ActiveTab =
+  | "models"
+  | "model"
+  | "techniques"
+  | "dataset"
+  | "mcp"
+  | "mcp_harness"
+  | "distill"
+  | "distillation"
+  | "pruning"
+  | "moe"
+  | "moe_merge"
+  | "gguf"
+  | "train"
+  | "training"
+  | "deploy"
+  | "ollama"
+  | "arena";