@elizaos/training 2.0.0-alpha.13 → 2.0.0-alpha.15

package/src/training/ModelFetcher.ts

@@ -4,8 +4,8 @@
  * Fetches trained RL models from the database for inference.
  */
 
-import { getTrainingDataAdapter } from '../adapter';
-import { logger } from '../utils/logger';
+import { getTrainingDataAdapter } from "../adapter";
+import { logger } from "../utils/logger";
 
 export interface ModelArtifact {
   version: string;
@@ -33,7 +33,7 @@ export async function getLatestRLModel(): Promise<ModelArtifact | null> {
   }
 
   // Skip models that aren't ready or deployed
-  if (model.status !== 'ready' && model.status !== 'deployed') {
+  if (model.status !== "ready" && model.status !== "deployed") {
     return null;
   }
 
@@ -41,23 +41,23 @@ export async function getLatestRLModel(): Promise<ModelArtifact | null> {
 
   if (!rlModelId || rlModelId.trim().length === 0) {
     logger.error(
-      'Model has no storagePath or modelId',
+      "Model has no storagePath or modelId",
       {
         modelId: model.modelId,
         storagePath: model.storagePath,
       },
-      'ModelFetcher'
+      "ModelFetcher",
     );
     return null;
   }
 
   if (!model.baseModel || model.baseModel.trim().length === 0) {
     logger.error(
-      'Model has no baseModel',
+      "Model has no baseModel",
       {
         modelId: model.modelId,
       },
-      'ModelFetcher'
+      "ModelFetcher",
     );
     return null;
   }

package/src/training/ModelSelectionService.ts

@@ -7,13 +7,13 @@
  * 3. Performance of previous models
  */
 
-import { getTrainingDataAdapter } from '../adapter';
-import { logger } from '../utils/logger';
+import { getTrainingDataAdapter } from "../adapter";
+import { logger } from "../utils/logger";
 
 export interface ModelSelectionResult {
   modelId: string;
   modelPath: string;
-  strategy: 'base' | 'continue' | 'force_first';
+  strategy: "base" | "continue" | "force_first";
   reason: string;
   metadata?: {
     bundleCount?: number;
@@ -32,7 +32,7 @@ export interface TrainingBundle {
 export class ModelSelectionService {
   /** Default base model - uses Qwen3-4B-128K (4B params, 128K context). Scale up via MODEL_TIER or AVAILABLE_VRAM_GB env vars */
   private readonly BASE_MODEL =
-    process.env.BASE_MODEL || 'unsloth/Qwen3-4B-128K';
+    process.env.BASE_MODEL || "unsloth/Qwen3-4B-128K";
   private readonly BUNDLE_THRESHOLD = 1000;
   private readonly MIN_BUNDLES_FOR_TRAINING = 100;
   private readonly MAX_TRAINING_EXAMPLES = 2000;
@@ -61,9 +61,9 @@ export class ModelSelectionService {
    */
   async selectBaseModel(): Promise<ModelSelectionResult> {
     logger.info(
-      'Selecting base model for training...',
+      "Selecting base model for training...",
       undefined,
-      'ModelSelectionService'
+      "ModelSelectionService",
     );
 
     // Count available training bundles (always fetch for accurate metrics)
@@ -74,15 +74,15 @@ export class ModelSelectionService {
 
     if (forceFirst) {
       logger.info(
-        'No models exist - forcing first model creation',
+        "No models exist - forcing first model creation",
         undefined,
-        'ModelSelectionService'
+        "ModelSelectionService",
       );
       return {
         modelId: this.BASE_MODEL,
         modelPath: this.BASE_MODEL,
-        strategy: 'force_first',
-        reason: 'No trained models exist - creating first model from base',
+        strategy: "force_first",
+        reason: "No trained models exist - creating first model from base",
         metadata: {
           baseModel: this.BASE_MODEL,
           bundleCount, // Use actual count, not 0
@@ -92,14 +92,14 @@ export class ModelSelectionService {
     logger.info(
       `Found ${bundleCount} training bundles`,
       undefined,
-      'ModelSelectionService'
+      "ModelSelectionService",
     );
 
     // Not enough data yet
     if (bundleCount < this.MIN_BUNDLES_FOR_TRAINING) {
       throw new Error(
         `Insufficient training data: ${bundleCount} bundles ` +
-          `(need ${this.MIN_BUNDLES_FOR_TRAINING} minimum)`
+          `(need ${this.MIN_BUNDLES_FOR_TRAINING} minimum)`,
       );
     }
 
@@ -108,12 +108,12 @@ export class ModelSelectionService {
       logger.info(
         `Bundle count ${bundleCount} < ${this.BUNDLE_THRESHOLD} - using base model`,
         undefined,
-        'ModelSelectionService'
+        "ModelSelectionService",
       );
       return {
         modelId: this.BASE_MODEL,
         modelPath: this.BASE_MODEL,
-        strategy: 'base',
+        strategy: "base",
         reason: `Training from base model (${bundleCount} bundles < ${this.BUNDLE_THRESHOLD} threshold)`,
         metadata: {
           bundleCount,
@@ -127,15 +127,15 @@ export class ModelSelectionService {
 
     if (!bestModel) {
       logger.warn(
-        'No best model found despite bundle threshold - using base model',
+        "No best model found despite bundle threshold - using base model",
         undefined,
-        'ModelSelectionService'
+        "ModelSelectionService",
       );
       return {
         modelId: this.BASE_MODEL,
         modelPath: this.BASE_MODEL,
-        strategy: 'base',
-        reason: 'No previous models available - using base model',
+        strategy: "base",
+        reason: "No previous models available - using base model",
         metadata: {
           bundleCount,
           baseModel: this.BASE_MODEL,
@@ -149,7 +149,7 @@ export class ModelSelectionService {
         bestModelId: bestModel.modelId,
         bestScore: bestModel.benchmarkScore,
       },
-      'ModelSelectionService'
+      "ModelSelectionService",
     );
 
     // Use storagePath for model path (e.g., HuggingFace URL)
@@ -158,8 +158,8 @@ export class ModelSelectionService {
     return {
       modelId: bestModel.modelId,
       modelPath: modelStoragePath,
-      strategy: 'continue',
-      reason: `Continuing from best model (score: ${bestModel.benchmarkScore?.toFixed(3) || 'N/A'})`,
+      strategy: "continue",
+      reason: `Continuing from best model (score: ${bestModel.benchmarkScore?.toFixed(3) || "N/A"})`,
       metadata: {
         bundleCount,
         bestModelScore: bestModel.benchmarkScore || undefined,
@@ -185,22 +185,22 @@ export class ModelSelectionService {
 
     if (!model) {
       logger.warn(
-        'No benchmarked models found',
+        "No benchmarked models found",
         undefined,
-        'ModelSelectionService'
+        "ModelSelectionService",
       );
       return null;
     }
 
     logger.info(
-      'Found best performing model',
+      "Found best performing model",
       {
         modelId: model.modelId,
         version: model.version,
         benchmarkScore: model.benchmarkScore,
         avgReward: model.avgReward,
       },
-      'ModelSelectionService'
+      "ModelSelectionService",
     );
 
     return model;
@@ -278,13 +278,13 @@ export class ModelSelectionService {
    */
   async getTrainingTrajectories(limit?: number | null) {
     const result = await getTrainingDataAdapter().getTrainingTrajectories(
-      limit ?? undefined
+      limit ?? undefined,
     );
 
     logger.info(
       `Retrieved ${result.length} trajectories for training`,
       { limit, available: result.length },
-      'ModelSelectionService'
+      "ModelSelectionService",
     );
 
     return result;
@@ -316,15 +316,15 @@ export class ModelSelectionService {
     const trainedModelCount = await this.countTrainedModels();
     const bestModel = await this.getBestPerformingModel();
 
-    let recommendation = '';
+    let recommendation = "";
     if (trainedModelCount === 0) {
-      recommendation = 'Force first model creation';
+      recommendation = "Force first model creation";
     } else if (bundleCount < this.MIN_BUNDLES_FOR_TRAINING) {
-      recommendation = 'Not ready - need more data';
+      recommendation = "Not ready - need more data";
     } else if (bundleCount < this.BUNDLE_THRESHOLD) {
-      recommendation = 'Train from base model';
+      recommendation = "Train from base model";
     } else {
-      recommendation = 'Train from best performing model';
+      recommendation = "Train from best performing model";
     }
 
     return {

package/src/training/ModelUsageVerifier.ts

@@ -5,14 +5,14 @@
  * Provides assertions and logging for model usage verification.
  */
 
-import { getTrainingDataAdapter, getLlmLogAdapter } from '../adapter';
-import type { IAgentRuntimeLike } from '../dependencies';
-import { logger } from '../utils/logger';
+import { getLlmLogAdapter, getTrainingDataAdapter } from "../adapter";
+import type { IAgentRuntimeLike } from "../dependencies";
+import { logger } from "../utils/logger";
 
 export interface ModelUsageStats {
   agentId: string;
   modelUsed: string;
-  modelSource: 'groq' | 'claude' | 'openai' | 'unknown';
+  modelSource: "groq" | "claude" | "openai" | "unknown";
   inferenceCount: number;
 }
 
@@ -36,7 +36,7 @@ export class ModelUsageVerifier {
    */
   static async verifyAgentModelUsage(
     agentUserId: string,
-    runtime: IAgentRuntimeLike
+    runtime: IAgentRuntimeLike,
   ): Promise<ModelUsageStats> {
     const character = (runtime as Record<string, unknown>).character as
       | { settings?: Record<string, unknown> }
@@ -45,30 +45,31 @@ export class ModelUsageVerifier {
 
     // Check for different model providers
     const groqModel = String(
-      settings?.GROQ_LARGE_MODEL || settings?.GROQ_SMALL_MODEL || ''
+      settings?.GROQ_LARGE_MODEL || settings?.GROQ_SMALL_MODEL || "",
     );
-    const claudeModel = String(settings?.CLAUDE_MODEL || '');
-    const openaiModel = String(settings?.OPENAI_MODEL || '');
+    const claudeModel = String(settings?.CLAUDE_MODEL || "");
+    const openaiModel = String(settings?.OPENAI_MODEL || "");
 
     let modelUsed: string;
-    let modelSource: 'groq' | 'claude' | 'openai' | 'unknown';
+    let modelSource: "groq" | "claude" | "openai" | "unknown";
 
     if (claudeModel) {
       modelUsed = claudeModel;
-      modelSource = 'claude';
+      modelSource = "claude";
     } else if (openaiModel) {
       modelUsed = openaiModel;
-      modelSource = 'openai';
+      modelSource = "openai";
     } else if (groqModel) {
       modelUsed = groqModel;
-      modelSource = 'groq';
+      modelSource = "groq";
     } else {
-      modelUsed = 'unknown';
-      modelSource = 'unknown';
+      modelUsed = "unknown";
+      modelSource = "unknown";
     }
 
     // Count inferences from logs (using trajectoryId)
-    const trajectoryIds = await getTrainingDataAdapter().getTrajectoryIdsByAgent(agentUserId);
+    const trajectoryIds =
+      await getTrainingDataAdapter().getTrajectoryIdsByAgent(agentUserId);
 
     const twentyFourHoursAgo = new Date(Date.now() - 24 * 60 * 60 * 1000);
 
@@ -77,7 +78,7 @@ export class ModelUsageVerifier {
     if (llmAdapter && trajectoryIds.length > 0) {
       inferenceCount = await llmAdapter.countRecentLLMCalls(
         trajectoryIds,
-        twentyFourHoursAgo
+        twentyFourHoursAgo,
       );
     }
 
@@ -94,7 +95,7 @@ export class ModelUsageVerifier {
    */
   static async verifyMultipleAgents(
     agentUserIds: string[],
-    runtimes: Map<string, IAgentRuntimeLike>
+    runtimes: Map<string, IAgentRuntimeLike>,
   ): Promise<VerificationResult> {
     const details: ModelUsageStats[] = [];
     const errors: string[] = [];
@@ -106,7 +107,10 @@ export class ModelUsageVerifier {
         continue;
       }
 
-      const stats = await this.verifyAgentModelUsage(agentId, runtime);
+      const stats = await ModelUsageVerifier.verifyAgentModelUsage(
+        agentId,
+        runtime,
+      );
       details.push(stats);
     }
 
@@ -123,25 +127,28 @@ export class ModelUsageVerifier {
    */
   static async assertModelUsage(
     agentUserId: string,
-    runtime: IAgentRuntimeLike
+    runtime: IAgentRuntimeLike,
   ): Promise<void> {
-    const stats = await this.verifyAgentModelUsage(agentUserId, runtime);
+    const stats = await ModelUsageVerifier.verifyAgentModelUsage(
+      agentUserId,
+      runtime,
+    );
 
-    if (stats.modelSource === 'unknown') {
+    if (stats.modelSource === "unknown") {
       throw new Error(
         `Agent ${agentUserId} has no configured model. ` +
-          `Using: ${stats.modelUsed}`
+          `Using: ${stats.modelUsed}`,
       );
     }
 
     logger.info(
-      'Model usage verified',
+      "Model usage verified",
       {
         agentId: agentUserId,
         model: stats.modelUsed,
         source: stats.modelSource,
       },
-      'ModelUsageVerifier'
+      "ModelUsageVerifier",
     );
   }
 

package/src/training/MultiModelOrchestrator.ts

@@ -11,7 +11,7 @@
  * - Real vLLM/OpenAI-compatible API integration
  */
 
-import { logger } from '../utils/logger';
+import { logger } from "../utils/logger";
 import {
   getModelForArchetype as getArchetypeModel,
   getMultiModelConfig,
@@ -20,7 +20,7 @@ import {
   type ModelTier,
   type MultiModelConfig,
   type QuantizationMode,
-} from './RLModelConfig';
+} from "./RLModelConfig";
 
 /**
  * Loaded model state
@@ -107,9 +107,9 @@ export class MultiModelOrchestrator {
 
   constructor(config: OrchestratorConfig) {
     this.config = {
-      vllmBaseUrl: process.env.VLLM_BASE_URL || 'http://localhost:9001',
+      vllmBaseUrl: process.env.VLLM_BASE_URL || "http://localhost:9001",
       fallbackApiUrl:
-        process.env.GROQ_API_URL || 'https://api.groq.com/openai/v1',
+        process.env.GROQ_API_URL || "https://api.groq.com/openai/v1",
       fallbackApiKey: process.env.GROQ_API_KEY,
       inferenceTimeoutMs: 30000,
       ...config,
@@ -117,7 +117,7 @@ export class MultiModelOrchestrator {
     this.multiModelConfig = getMultiModelConfig(config.availableVramGb);
 
     logger.info(
-      'MultiModelOrchestrator initialized',
+      "MultiModelOrchestrator initialized",
       {
         availableVram: `${config.availableVramGb}GB`,
         maxConcurrentModels: this.multiModelConfig.maxConcurrentModels,
@@ -126,7 +126,7 @@ export class MultiModelOrchestrator {
         vllmUrl: this.config.vllmBaseUrl,
         hasFallback: !!this.config.fallbackApiKey,
       },
-      'MultiModelOrchestrator'
+      "MultiModelOrchestrator",
     );
   }
 
@@ -150,9 +150,9 @@ export class MultiModelOrchestrator {
 
       if (this.vllmAvailable) {
         logger.info(
-          'vLLM server is available',
+          "vLLM server is available",
           { url: this.config.vllmBaseUrl },
-          'MultiModelOrchestrator'
+          "MultiModelOrchestrator",
         );
       }
 
@@ -161,9 +161,9 @@ export class MultiModelOrchestrator {
       clearTimeout(timeout);
       this.vllmAvailable = false;
       logger.warn(
-        'vLLM server not available, will use fallback',
+        "vLLM server not available, will use fallback",
         { url: this.config.vllmBaseUrl },
-        'MultiModelOrchestrator'
+        "MultiModelOrchestrator",
       );
       return false;
     }
@@ -187,7 +187,7 @@ export class MultiModelOrchestrator {
         quantization: this.config.defaultQuantization,
         vramGb: getVramRequirement(
           this.config.defaultTier,
-          this.config.defaultQuantization
+          this.config.defaultQuantization,
         ),
       };
     }
@@ -237,7 +237,7 @@ export class MultiModelOrchestrator {
             freedVram: `${model.vramUsageGb}GB`,
             currentUsage: `${this.currentVramUsageGb}GB`,
           },
-          'MultiModelOrchestrator'
+          "MultiModelOrchestrator",
         );
       }
     }
@@ -263,7 +263,7 @@ export class MultiModelOrchestrator {
     if (!this.canLoadModel(modelInfo.vramGb)) {
       throw new Error(
         `Cannot load model for ${archetype}: insufficient VRAM. ` +
-          `Required: ${modelInfo.vramGb}GB, Available: ${this.config.availableVramGb - this.currentVramUsageGb}GB`
+          `Required: ${modelInfo.vramGb}GB, Available: ${this.config.availableVramGb - this.currentVramUsageGb}GB`,
       );
     }
 
@@ -288,7 +288,7 @@ export class MultiModelOrchestrator {
         totalVramUsed: `${this.currentVramUsageGb}GB`,
         modelsLoaded: this.loadedModels.size,
       },
-      'MultiModelOrchestrator'
+      "MultiModelOrchestrator",
     );
 
     return loadedModel;
@@ -302,32 +302,32 @@ export class MultiModelOrchestrator {
     prompt: string,
     systemPrompt: string,
     maxTokens: number,
-    temperature: number
+    temperature: number,
   ): Promise<CompletionResponse> {
     const controller = new AbortController();
     const timeout = setTimeout(
       () => controller.abort(),
-      this.config.inferenceTimeoutMs
+      this.config.inferenceTimeoutMs,
     );
 
     const response = await fetch(
       `${this.config.vllmBaseUrl}/v1/chat/completions`,
       {
-        method: 'POST',
+        method: "POST",
         headers: {
-          'Content-Type': 'application/json',
+          "Content-Type": "application/json",
         },
         body: JSON.stringify({
           model: modelId,
           messages: [
-            { role: 'system', content: systemPrompt },
-            { role: 'user', content: prompt },
+            { role: "system", content: systemPrompt },
+            { role: "user", content: prompt },
           ],
           max_tokens: maxTokens,
           temperature,
         }),
         signal: controller.signal,
-      }
+      },
     );
 
     clearTimeout(timeout);
@@ -347,42 +347,42 @@ export class MultiModelOrchestrator {
     prompt: string,
     systemPrompt: string,
     maxTokens: number,
-    temperature: number
+    temperature: number,
   ): Promise<CompletionResponse> {
     if (!this.config.fallbackApiKey) {
       throw new Error(
-        'No fallback API key configured. Set GROQ_API_KEY environment variable.'
+        "No fallback API key configured. Set GROQ_API_KEY environment variable.",
       );
     }
 
     const controller = new AbortController();
     const timeout = setTimeout(
       () => controller.abort(),
-      this.config.inferenceTimeoutMs
+      this.config.inferenceTimeoutMs,
     );
 
     // Use a fast model for fallback
-    const fallbackModel = 'llama-3.1-8b-instant';
+    const fallbackModel = "llama-3.1-8b-instant";
 
     const response = await fetch(
       `${this.config.fallbackApiUrl}/chat/completions`,
       {
-        method: 'POST',
+        method: "POST",
         headers: {
-          'Content-Type': 'application/json',
+          "Content-Type": "application/json",
           Authorization: `Bearer ${this.config.fallbackApiKey}`,
         },
         body: JSON.stringify({
           model: fallbackModel,
           messages: [
-            { role: 'system', content: systemPrompt },
-            { role: 'user', content: prompt },
+            { role: "system", content: systemPrompt },
+            { role: "user", content: prompt },
           ],
           max_tokens: maxTokens,
           temperature,
         }),
         signal: controller.signal,
-      }
+      },
     );
 
     clearTimeout(timeout);
@@ -390,7 +390,7 @@ export class MultiModelOrchestrator {
     if (!response.ok) {
       const error = await response.text();
       throw new Error(
-        `Fallback API request failed: ${response.status} - ${error}`
+        `Fallback API request failed: ${response.status} - ${error}`,
       );
     }
 
@@ -401,7 +401,7 @@ export class MultiModelOrchestrator {
    * Run inference for an archetype
    */
   async inference(
-    request: ModelInferenceRequest
+    request: ModelInferenceRequest,
   ): Promise<ModelInferenceResult> {
     const startTime = Date.now();
 
@@ -427,7 +427,7 @@ export class MultiModelOrchestrator {
           request.prompt,
           systemPrompt,
           maxTokens,
-          temperature
+          temperature,
         );
       } else {
         // Fall back to Groq/OpenAI
@@ -435,12 +435,12 @@ export class MultiModelOrchestrator {
           request.prompt,
           systemPrompt,
           maxTokens,
-          temperature
+          temperature,
         );
       }
 
       const latencyMs = Date.now() - startTime;
-      const response = completion.choices[0]?.message.content || '';
+      const response = completion.choices[0]?.message.content || "";
       const tokensGenerated = completion.usage?.completion_tokens || 0;
 
       logger.debug(
@@ -451,7 +451,7 @@ export class MultiModelOrchestrator {
           tokensGenerated,
           usedVllm: vllmAvailable,
         },
-        'MultiModelOrchestrator'
+        "MultiModelOrchestrator",
       );
 
       return {
@@ -469,12 +469,12 @@ export class MultiModelOrchestrator {
       logger.error(
         `Inference failed for ${request.archetype}`,
         { error: errorMessage, latencyMs },
-        'MultiModelOrchestrator'
+        "MultiModelOrchestrator",
       );
 
       return {
         archetype: request.archetype,
-        response: '',
+        response: "",
         modelId: model.modelId,
         latencyMs,
         tokensGenerated: 0,
@@ -487,7 +487,7 @@ export class MultiModelOrchestrator {
    * Batch inference for multiple archetypes
    */
   async batchInference(
-    requests: ModelInferenceRequest[]
+    requests: ModelInferenceRequest[],
   ): Promise<ModelInferenceResult[]> {
     // Group requests by archetype for efficient batching
     const byArchetype = new Map<string, ModelInferenceRequest[]>();
@@ -509,7 +509,7 @@ export class MultiModelOrchestrator {
       for (let i = 0; i < archetypeRequests.length; i += batchSize) {
         const batch = archetypeRequests.slice(i, i + batchSize);
         const batchResults = await Promise.all(
-          batch.map((req) => this.inference(req))
+          batch.map((req) => this.inference(req)),
         );
         results.push(...batchResults);
       }
@@ -555,7 +555,7 @@ export class MultiModelOrchestrator {
   unloadAll(): void {
     this.loadedModels.clear();
     this.currentVramUsageGb = 0;
-    logger.info('Unloaded all models', {}, 'MultiModelOrchestrator');
+    logger.info("Unloaded all models", {}, "MultiModelOrchestrator");
   }
 
   /**
@@ -570,11 +570,11 @@ export class MultiModelOrchestrator {
  * Create a multi-model orchestrator with sensible defaults for RTX 5090 (16GB)
  */
 export function createMultiModelOrchestrator(
-  vramGb = 16
+  vramGb = 16,
 ): MultiModelOrchestrator {
   return new MultiModelOrchestrator({
     availableVramGb: vramGb,
-    defaultTier: 'small',
-    defaultQuantization: '4bit',
+    defaultTier: "small",
+    defaultQuantization: "4bit",
   });
 }