@elizaos/plugin-training 2.0.3-beta.5 → 2.0.3-beta.7

package/dist/dspy/optimizers/dspy-bootstrap-fewshot.d.ts

@@ -0,0 +1,23 @@
+/**
+ * DSPy-style BootstrapFewshot optimizer.
+ *
+ * Selects top-K examples from the dataset (ranked by reward, then by metric
+ * score against the baseline) and emits them as the demonstration set
+ * attached to a Predict module. Unlike the legacy `bootstrap-fewshot.ts`
+ * which renders demos as text appended to a baseline prompt, this version
+ * keeps demonstrations as a typed `Example[]` so they round-trip through
+ * Signature.render() at run time.
+ */
+import type { DspyOptimizerInput, DspyOptimizerResult } from "./types.js";
+export interface DspyBootstrapFewshotOptions {
+    /** Demonstrations to keep. Defaults to 5. */
+    k?: number;
+    /**
+     * When true, rank examples by their per-example metric score against the
+     * baseline signature instead of by the recorded `reward`. Slower (one LM
+     * call per example) but works when rewards are missing.
+     */
+    rankByMetric?: boolean;
+}
+export declare function runDspyBootstrapFewshot(input: DspyOptimizerInput, options?: DspyBootstrapFewshotOptions): Promise<DspyOptimizerResult>;
+//# sourceMappingURL=dspy-bootstrap-fewshot.d.ts.map

package/dist/dspy/optimizers/dspy-bootstrap-fewshot.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"dspy-bootstrap-fewshot.d.ts","sourceRoot":"","sources":["../../../src/dspy/optimizers/dspy-bootstrap-fewshot.ts"],"names":[],"mappings":"AAAA;;;;;;;;;GASG;AAIH,OAAO,KAAK,EACV,kBAAkB,EAClB,mBAAmB,EAEpB,MAAM,YAAY,CAAC;AAEpB,MAAM,WAAW,2BAA2B;IAC1C,6CAA6C;IAC7C,CAAC,CAAC,EAAE,MAAM,CAAC;IACX;;;;OAIG;IACH,YAAY,CAAC,EAAE,OAAO,CAAC;CACxB;AAED,wBAAsB,uBAAuB,CAC3C,KAAK,EAAE,kBAAkB,EACzB,OAAO,GAAE,2BAAgC,GACxC,OAAO,CAAC,mBAAmB,CAAC,CAgC9B"}

package/dist/dspy/optimizers/dspy-bootstrap-fewshot.js

@@ -0,0 +1,85 @@
+import { Predict } from "../predict.js";
+async function runDspyBootstrapFewshot(input, options = {}) {
+  const k = Math.max(1, options.k ?? 5);
+  const lineage = [];
+  const baselineScore = await evaluate(input, []);
+  lineage.push({
+    round: 0,
+    variant: 0,
+    score: baselineScore,
+    notes: "baseline"
+  });
+  const ranked = await rankExamples(input, options);
+  const demos = ranked.slice(0, Math.min(k, ranked.length));
+  const optimizedScore = await evaluate(input, demos);
+  lineage.push({
+    round: 1,
+    variant: 1,
+    score: optimizedScore,
+    notes: `injected ${demos.length} demonstrations`
+  });
+  return {
+    optimizer: "dspy-bootstrap-fewshot",
+    signature: input.signature,
+    instructions: input.signature.spec.instructions,
+    demonstrations: demos,
+    score: optimizedScore,
+    baselineScore,
+    lineage
+  };
+}
+async function evaluate(input, demonstrations) {
+  if (input.dataset.length === 0) return 0;
+  const predict = new Predict({
+    signature: input.signature,
+    lm: input.lm,
+    demonstrations
+  });
+  let total = 0;
+  let counted = 0;
+  for (const example of input.dataset) {
+    try {
+      const { output } = await predict.forward(example.inputs);
+      total += input.metric(output, example.outputs);
+      counted += 1;
+    } catch {
+      counted += 1;
+    }
+  }
+  return counted === 0 ? 0 : total / counted;
+}
+async function rankExamples(input, options) {
+  if (options.rankByMetric) {
+    const scored = [];
+    for (const example of input.dataset) {
+      const predict = new Predict({
+        signature: input.signature,
+        lm: input.lm,
+        demonstrations: []
+      });
+      let score = 0;
+      try {
+        const { output } = await predict.forward(example.inputs);
+        score = input.metric(output, example.outputs);
+      } catch {
+        score = 0;
+      }
+      scored.push({ example, score });
+    }
+    scored.sort(
+      (a, b) => b.score - a.score || (b.example.reward ?? 0) - (a.example.reward ?? 0)
+    );
+    return scored.map((s) => s.example);
+  }
+  const ordered = input.dataset.map((example, index) => ({
+    example,
+    index,
+    reward: example.reward ?? 0
+  }));
+  ordered.sort((a, b) => b.reward - a.reward || a.index - b.index);
+  return ordered.map((entry) => entry.example);
+}
+export {
+  runDspyBootstrapFewshot
+};
+//# sourceMappingURL=dspy-bootstrap-fewshot.js.map

package/dist/dspy/optimizers/dspy-bootstrap-fewshot.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../../../src/dspy/optimizers/dspy-bootstrap-fewshot.ts"],"sourcesContent":["/**\n * DSPy-style BootstrapFewshot optimizer.\n *\n * Selects top-K examples from the dataset (ranked by reward, then by metric\n * score against the baseline) and emits them as the demonstration set\n * attached to a Predict module. Unlike the legacy `bootstrap-fewshot.ts`\n * which renders demos as text appended to a baseline prompt, this version\n * keeps demonstrations as a typed `Example[]` so they round-trip through\n * Signature.render() at run time.\n */\n\nimport type { Example } from \"../examples.js\";\nimport { Predict } from \"../predict.js\";\nimport type {\n  DspyOptimizerInput,\n  DspyOptimizerResult,\n  OptimizerLineageEntry,\n} from \"./types.js\";\n\nexport interface DspyBootstrapFewshotOptions {\n  /** Demonstrations to keep. Defaults to 5. */\n  k?: number;\n  /**\n   * When true, rank examples by their per-example metric score against the\n   * baseline signature instead of by the recorded `reward`. Slower (one LM\n   * call per example) but works when rewards are missing.\n   */\n  rankByMetric?: boolean;\n}\n\nexport async function runDspyBootstrapFewshot(\n  input: DspyOptimizerInput,\n  options: DspyBootstrapFewshotOptions = {},\n): Promise<DspyOptimizerResult> {\n  const k = Math.max(1, options.k ?? 5);\n  const lineage: OptimizerLineageEntry[] = [];\n\n  const baselineScore = await evaluate(input, []);\n  lineage.push({\n    round: 0,\n    variant: 0,\n    score: baselineScore,\n    notes: \"baseline\",\n  });\n\n  const ranked = await rankExamples(input, options);\n  const demos = ranked.slice(0, Math.min(k, ranked.length));\n\n  const optimizedScore = await evaluate(input, demos);\n  lineage.push({\n    round: 1,\n    variant: 1,\n    score: optimizedScore,\n    notes: `injected ${demos.length} demonstrations`,\n  });\n\n  return {\n    optimizer: \"dspy-bootstrap-fewshot\",\n    signature: input.signature,\n    instructions: input.signature.spec.instructions,\n    demonstrations: demos,\n    score: optimizedScore,\n    baselineScore,\n    lineage,\n  };\n}\n\nasync function evaluate(\n  input: DspyOptimizerInput,\n  demonstrations: Example[],\n): Promise<number> {\n  if (input.dataset.length === 0) return 0;\n  const predict = new Predict({\n    signature: input.signature,\n    lm: input.lm,\n    demonstrations,\n  });\n  let total = 0;\n  let counted = 0;\n  for (const example of input.dataset) {\n    try {\n      const { output } = await predict.forward(example.inputs);\n      total += input.metric(output, example.outputs);\n      counted += 1;\n    } catch {\n      // Parse failure scores 0 — the optimizer must learn to produce\n      // signature-compatible output. Avoid burying the failure with `?? 0`\n      // at the metric layer.\n      counted += 1;\n    }\n  }\n  return counted === 0 ? 0 : total / counted;\n}\n\nasync function rankExamples(\n  input: DspyOptimizerInput,\n  options: DspyBootstrapFewshotOptions,\n): Promise<Example[]> {\n  if (options.rankByMetric) {\n    const scored: Array<{ example: Example; score: number }> = [];\n    for (const example of input.dataset) {\n      const predict = new Predict({\n        signature: input.signature,\n        lm: input.lm,\n        demonstrations: [],\n      });\n      let score = 0;\n      try {\n        const { output } = await predict.forward(example.inputs);\n        score = input.metric(output, example.outputs);\n      } catch {\n        score = 0;\n      }\n      scored.push({ example, score });\n    }\n    scored.sort(\n      (a, b) =>\n        b.score - a.score || (b.example.reward ?? 0) - (a.example.reward ?? 0),\n    );\n    return scored.map((s) => s.example);\n  }\n  const ordered = input.dataset.map((example, index) => ({\n    example,\n    index,\n    reward: example.reward ?? 0,\n  }));\n  ordered.sort((a, b) => b.reward - a.reward || a.index - b.index);\n  return ordered.map((entry) => entry.example);\n}\n"],"mappings":"AAYA,SAAS,eAAe;AAkBxB,eAAsB,wBACpB,OACA,UAAuC,CAAC,GACV;AAC9B,QAAM,IAAI,KAAK,IAAI,GAAG,QAAQ,KAAK,CAAC;AACpC,QAAM,UAAmC,CAAC;AAE1C,QAAM,gBAAgB,MAAM,SAAS,OAAO,CAAC,CAAC;AAC9C,UAAQ,KAAK;AAAA,IACX,OAAO;AAAA,IACP,SAAS;AAAA,IACT,OAAO;AAAA,IACP,OAAO;AAAA,EACT,CAAC;AAED,QAAM,SAAS,MAAM,aAAa,OAAO,OAAO;AAChD,QAAM,QAAQ,OAAO,MAAM,GAAG,KAAK,IAAI,GAAG,OAAO,MAAM,CAAC;AAExD,QAAM,iBAAiB,MAAM,SAAS,OAAO,KAAK;AAClD,UAAQ,KAAK;AAAA,IACX,OAAO;AAAA,IACP,SAAS;AAAA,IACT,OAAO;AAAA,IACP,OAAO,YAAY,MAAM,MAAM;AAAA,EACjC,CAAC;AAED,SAAO;AAAA,IACL,WAAW;AAAA,IACX,WAAW,MAAM;AAAA,IACjB,cAAc,MAAM,UAAU,KAAK;AAAA,IACnC,gBAAgB;AAAA,IAChB,OAAO;AAAA,IACP;AAAA,IACA;AAAA,EACF;AACF;AAEA,eAAe,SACb,OACA,gBACiB;AACjB,MAAI,MAAM,QAAQ,WAAW,EAAG,QAAO;AACvC,QAAM,UAAU,IAAI,QAAQ;AAAA,IAC1B,WAAW,MAAM;AAAA,IACjB,IAAI,MAAM;AAAA,IACV;AAAA,EACF,CAAC;AACD,MAAI,QAAQ;AACZ,MAAI,UAAU;AACd,aAAW,WAAW,MAAM,SAAS;AACnC,QAAI;AACF,YAAM,EAAE,OAAO,IAAI,MAAM,QAAQ,QAAQ,QAAQ,MAAM;AACvD,eAAS,MAAM,OAAO,QAAQ,QAAQ,OAAO;AAC7C,iBAAW;AAAA,IACb,QAAQ;AAIN,iBAAW;AAAA,IACb;AAAA,EACF;AACA,SAAO,YAAY,IAAI,IAAI,QAAQ;AACrC;AAEA,eAAe,aACb,OACA,SACoB;AACpB,MAAI,QAAQ,cAAc;AACxB,UAAM,SAAqD,CAAC;AAC5D,eAAW,WAAW,MAAM,SAAS;AACnC,YAAM,UAAU,IAAI,QAAQ;AAAA,QAC1B,WAAW,MAAM;AAAA,QACjB,IAAI,MAAM;AAAA,QACV,gBAAgB,CAAC;AAAA,MACnB,CAAC;AACD,UAAI,QAAQ;AACZ,UAAI;AACF,cAAM,EAAE,OAAO,IAAI,MAAM,QAAQ,QAAQ,QAAQ,MAAM;AACvD,gBAAQ,MAAM,OAAO,QAAQ,QAAQ,OAAO;AAAA,MAC9C,QAAQ;AACN,gBAAQ;AAAA,MACV;AACA,aAAO,KAAK,EAAE,SAAS,MAAM,CAAC;AAAA,IAChC;AACA,WAAO;AAAA,MACL,CAAC,GAAG,MACF,EAAE,QAAQ,EAAE,UAAU,EAAE,QAAQ,UAAU,MAAM,EAAE,QAAQ,UAAU;AAAA,IACxE;AACA,WAAO,OAAO,IAAI,CAAC,MAAM,EAAE,OAAO;AAAA,EACpC;AACA,QAAM,UAAU,MAAM,QAAQ,IAAI,CAAC,SAAS,WAAW;AAAA,IACrD;AAAA,IACA;AAAA,IACA,QAAQ,QAAQ,UAAU;AAAA,EAC5B,EAAE;AACF,UAAQ,KAAK,CAAC,GAAG,MAAM,EAAE,SAAS,EAAE,UAAU,EAAE,QAAQ,EAAE,KAAK;AAC/D,SAAO,QAAQ,IAAI,CAAC,UAAU,MAAM,OAAO;AAC7C;","names":[]}

package/dist/dspy/optimizers/dspy-copro.d.ts

@@ -0,0 +1,29 @@
+/**
+ * DSPy-style COPRO (Coordinate ascent over Prompt instructions) optimizer.
+ *
+ * Loop:
+ *   for round in 1..depth:
+ *     1. Propose N instruction variants via the teacher LM.
+ *     2. Score each variant on a held-out subset of the dataset.
+ *     3. Keep the highest-scoring variant as the new baseline for the next
+ *        round (greedy coordinate ascent).
+ *
+ * Returns the best (instructions, demonstrations=[]) pair observed. The
+ * caller can then re-train demonstrations on top of the winning instructions
+ * via BootstrapFewshot, which is what MIPRO does internally.
+ */
+import type { DspyOptimizerInput, DspyOptimizerResult } from "./types.js";
+export interface DspyCoproOptions {
+    /** Variants to propose per round. Defaults to 6. */
+    variants?: number;
+    /** Rounds of coordinate ascent. Defaults to 3. */
+    depth?: number;
+    /** Eval-set size per variant. Defaults to all examples. */
+    evalSubset?: number;
+    /** Teacher temperature for instruction proposal. Defaults to 0.8. */
+    teacherTemperature?: number;
+    /** Deterministic RNG for evalSubset sampling. Defaults to Math.random. */
+    rng?: () => number;
+}
+export declare function runDspyCopro(input: DspyOptimizerInput, options?: DspyCoproOptions): Promise<DspyOptimizerResult>;
+//# sourceMappingURL=dspy-copro.d.ts.map

package/dist/dspy/optimizers/dspy-copro.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"dspy-copro.d.ts","sourceRoot":"","sources":["../../../src/dspy/optimizers/dspy-copro.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;GAaG;AAMH,OAAO,KAAK,EACV,kBAAkB,EAClB,mBAAmB,EAGpB,MAAM,YAAY,CAAC;AAEpB,MAAM,WAAW,gBAAgB;IAC/B,oDAAoD;IACpD,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,kDAAkD;IAClD,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,2DAA2D;IAC3D,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,qEAAqE;IACrE,kBAAkB,CAAC,EAAE,MAAM,CAAC;IAC5B,0EAA0E;IAC1E,GAAG,CAAC,EAAE,MAAM,MAAM,CAAC;CACpB;AASD,wBAAsB,YAAY,CAChC,KAAK,EAAE,kBAAkB,EACzB,OAAO,GAAE,gBAAqB,GAC7B,OAAO,CAAC,mBAAmB,CAAC,CAmF9B"}

package/dist/dspy/optimizers/dspy-copro.js

@@ -0,0 +1,141 @@
+import { Predict } from "../predict.js";
+import { Signature } from "../signature.js";
+const PROPOSAL_SYSTEM = `You are rewriting the INSTRUCTIONS field of a task-prompt signature. The signature declares input fields, output fields, and a natural-language INSTRUCTIONS body. Your job is to produce a new INSTRUCTIONS body that makes a downstream language model perform the task more reliably.
+
+Hard constraints:
+- Preserve the semantic contract \u2014 the new instructions must still describe the same task and reference the same output fields.
+- Keep the rewrite concise (no preamble, no markdown headers, no role-play framing).
+- Output ONLY the rewritten instructions body \u2014 no fences, no commentary.`;
+async function runDspyCopro(input, options = {}) {
+  const variants = Math.max(1, options.variants ?? 6);
+  const depth = Math.max(1, options.depth ?? 3);
+  const teacher = input.teacher ?? input.lm;
+  const teacherTemperature = options.teacherTemperature ?? 0.8;
+  const rng = options.rng ?? Math.random;
+  const lineage = [];
+  const heldOut = typeof options.evalSubset === "number" ? subsample(input.dataset, options.evalSubset, rng) : input.dataset;
+  const baselineInstructions = input.signature.spec.instructions;
+  const baselineScore = await scoreInstructions(
+    baselineInstructions,
+    input.signature,
+    heldOut,
+    input.lm,
+    input.metric
+  );
+  lineage.push({
+    round: 0,
+    variant: 0,
+    score: baselineScore,
+    notes: "baseline"
+  });
+  let bestInstructions = baselineInstructions;
+  let bestScore = baselineScore;
+  let currentInstructions = baselineInstructions;
+  for (let round = 1; round <= depth; round += 1) {
+    let roundBestInstructions = currentInstructions;
+    let roundBestScore = bestScore;
+    for (let variant = 1; variant <= variants; variant += 1) {
+      const candidate = await proposeInstructions(
+        teacher,
+        input.signature,
+        currentInstructions,
+        teacherTemperature
+      );
+      if (candidate.trim().length === 0) {
+        lineage.push({
+          round,
+          variant,
+          score: 0,
+          notes: "empty proposal \u2014 skipped"
+        });
+        continue;
+      }
+      const score = await scoreInstructions(
+        candidate,
+        input.signature,
+        heldOut,
+        input.lm,
+        input.metric
+      );
+      lineage.push({ round, variant, score });
+      if (score > roundBestScore) {
+        roundBestScore = score;
+        roundBestInstructions = candidate;
+      }
+    }
+    if (roundBestScore > bestScore) {
+      bestScore = roundBestScore;
+      bestInstructions = roundBestInstructions;
+    }
+    currentInstructions = roundBestInstructions;
+  }
+  return {
+    optimizer: "dspy-copro",
+    signature: new Signature({
+      ...input.signature.spec,
+      instructions: bestInstructions
+    }),
+    instructions: bestInstructions,
+    demonstrations: [],
+    score: bestScore,
+    baselineScore,
+    lineage
+  };
+}
+async function proposeInstructions(teacher, signature, current, temperature) {
+  const ioSummary = [
+    `signature name: ${signature.spec.name}`,
+    "input fields:",
+    ...signature.spec.inputs.map(
+      (f) => `- ${f.name} (${f.type}): ${f.description}`
+    ),
+    "output fields:",
+    ...signature.spec.outputs.map(
+      (f) => `- ${f.name} (${f.type}): ${f.description}`
+    ),
+    "",
+    "Current INSTRUCTIONS body:",
+    current
+  ].join("\n");
+  const result = await teacher.generate({
+    system: PROPOSAL_SYSTEM,
+    messages: [{ role: "user", content: ioSummary }],
+    temperature,
+    maxTokens: 1024
+  });
+  return result.text.trim().replace(/^```[a-z0-9_-]*\s*/i, "").replace(/\s*```$/i, "").trim();
+}
+async function scoreInstructions(instructions, signature, dataset, lm, metric) {
+  if (dataset.length === 0) return 0;
+  const predict = new Predict({
+    signature,
+    lm,
+    instructionsOverride: instructions
+  });
+  let total = 0;
+  for (const example of dataset) {
+    try {
+      const { output } = await predict.forward(example.inputs);
+      total += metric(output, example.outputs);
+    } catch {
+    }
+  }
+  return total / dataset.length;
+}
+function subsample(items, count, rng) {
+  if (count >= items.length) return [...items];
+  const indices = /* @__PURE__ */ new Set();
+  const out = [];
+  while (out.length < count) {
+    const idx = Math.floor(rng() * items.length);
+    if (indices.has(idx)) continue;
+    indices.add(idx);
+    const item = items[idx];
+    if (item !== void 0) out.push(item);
+  }
+  return out;
+}
+export {
+  runDspyCopro
+};
+//# sourceMappingURL=dspy-copro.js.map

package/dist/dspy/optimizers/dspy-copro.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../../../src/dspy/optimizers/dspy-copro.ts"],"sourcesContent":["/**\n * DSPy-style COPRO (Coordinate ascent over Prompt instructions) optimizer.\n *\n * Loop:\n *   for round in 1..depth:\n *     1. Propose N instruction variants via the teacher LM.\n *     2. Score each variant on a held-out subset of the dataset.\n *     3. Keep the highest-scoring variant as the new baseline for the next\n *        round (greedy coordinate ascent).\n *\n * Returns the best (instructions, demonstrations=[]) pair observed. The\n * caller can then re-train demonstrations on top of the winning instructions\n * via BootstrapFewshot, which is what MIPRO does internally.\n */\n\nimport type { Example } from \"../examples.js\";\nimport type { LanguageModelAdapter } from \"../lm-adapter.js\";\nimport { Predict } from \"../predict.js\";\nimport { Signature } from \"../signature.js\";\nimport type {\n  DspyOptimizerInput,\n  DspyOptimizerResult,\n  Metric,\n  OptimizerLineageEntry,\n} from \"./types.js\";\n\nexport interface DspyCoproOptions {\n  /** Variants to propose per round. Defaults to 6. */\n  variants?: number;\n  /** Rounds of coordinate ascent. Defaults to 3. */\n  depth?: number;\n  /** Eval-set size per variant. Defaults to all examples. */\n  evalSubset?: number;\n  /** Teacher temperature for instruction proposal. Defaults to 0.8. */\n  teacherTemperature?: number;\n  /** Deterministic RNG for evalSubset sampling. Defaults to Math.random. */\n  rng?: () => number;\n}\n\nconst PROPOSAL_SYSTEM = `You are rewriting the INSTRUCTIONS field of a task-prompt signature. The signature declares input fields, output fields, and a natural-language INSTRUCTIONS body. Your job is to produce a new INSTRUCTIONS body that makes a downstream language model perform the task more reliably.\n\nHard constraints:\n- Preserve the semantic contract — the new instructions must still describe the same task and reference the same output fields.\n- Keep the rewrite concise (no preamble, no markdown headers, no role-play framing).\n- Output ONLY the rewritten instructions body — no fences, no commentary.`;\n\nexport async function runDspyCopro(\n  input: DspyOptimizerInput,\n  options: DspyCoproOptions = {},\n): Promise<DspyOptimizerResult> {\n  const variants = Math.max(1, options.variants ?? 6);\n  const depth = Math.max(1, options.depth ?? 3);\n  const teacher = input.teacher ?? input.lm;\n  const teacherTemperature = options.teacherTemperature ?? 0.8;\n  const rng = options.rng ?? Math.random;\n  const lineage: OptimizerLineageEntry[] = [];\n\n  const heldOut =\n    typeof options.evalSubset === \"number\"\n      ? subsample(input.dataset, options.evalSubset, rng)\n      : input.dataset;\n\n  const baselineInstructions = input.signature.spec.instructions;\n  const baselineScore = await scoreInstructions(\n    baselineInstructions,\n    input.signature,\n    heldOut,\n    input.lm,\n    input.metric,\n  );\n  lineage.push({\n    round: 0,\n    variant: 0,\n    score: baselineScore,\n    notes: \"baseline\",\n  });\n\n  let bestInstructions = baselineInstructions;\n  let bestScore = baselineScore;\n  let currentInstructions = baselineInstructions;\n\n  for (let round = 1; round <= depth; round += 1) {\n    let roundBestInstructions = currentInstructions;\n    let roundBestScore = bestScore;\n    for (let variant = 1; variant <= variants; variant += 1) {\n      const candidate = await proposeInstructions(\n        teacher,\n        input.signature,\n        currentInstructions,\n        teacherTemperature,\n      );\n      if (candidate.trim().length === 0) {\n        lineage.push({\n          round,\n          variant,\n          score: 0,\n          notes: \"empty proposal — skipped\",\n        });\n        continue;\n      }\n      const score = await scoreInstructions(\n        candidate,\n        input.signature,\n        heldOut,\n        input.lm,\n        input.metric,\n      );\n      lineage.push({ round, variant, score });\n      if (score > roundBestScore) {\n        roundBestScore = score;\n        roundBestInstructions = candidate;\n      }\n    }\n    if (roundBestScore > bestScore) {\n      bestScore = roundBestScore;\n      bestInstructions = roundBestInstructions;\n    }\n    currentInstructions = roundBestInstructions;\n  }\n\n  return {\n    optimizer: \"dspy-copro\",\n    signature: new Signature({\n      ...input.signature.spec,\n      instructions: bestInstructions,\n    }),\n    instructions: bestInstructions,\n    demonstrations: [],\n    score: bestScore,\n    baselineScore,\n    lineage,\n  };\n}\n\nasync function proposeInstructions(\n  teacher: LanguageModelAdapter,\n  signature: import(\"../signature.js\").Signature,\n  current: string,\n  temperature: number,\n): Promise<string> {\n  const ioSummary = [\n    `signature name: ${signature.spec.name}`,\n    \"input fields:\",\n    ...signature.spec.inputs.map(\n      (f) => `- ${f.name} (${f.type}): ${f.description}`,\n    ),\n    \"output fields:\",\n    ...signature.spec.outputs.map(\n      (f) => `- ${f.name} (${f.type}): ${f.description}`,\n    ),\n    \"\",\n    \"Current INSTRUCTIONS body:\",\n    current,\n  ].join(\"\\n\");\n  const result = await teacher.generate({\n    system: PROPOSAL_SYSTEM,\n    messages: [{ role: \"user\", content: ioSummary }],\n    temperature,\n    maxTokens: 1024,\n  });\n  return result.text\n    .trim()\n    .replace(/^```[a-z0-9_-]*\\s*/i, \"\")\n    .replace(/\\s*```$/i, \"\")\n    .trim();\n}\n\nasync function scoreInstructions(\n  instructions: string,\n  signature: import(\"../signature.js\").Signature,\n  dataset: Example[],\n  lm: LanguageModelAdapter,\n  metric: Metric,\n): Promise<number> {\n  if (dataset.length === 0) return 0;\n  const predict = new Predict({\n    signature,\n    lm,\n    instructionsOverride: instructions,\n  });\n  let total = 0;\n  for (const example of dataset) {\n    try {\n      const { output } = await predict.forward(example.inputs);\n      total += metric(output, example.outputs);\n    } catch {\n      // Parse failure is a real failure for scoring. No silent fallback.\n    }\n  }\n  return total / dataset.length;\n}\n\nfunction subsample<T>(items: T[], count: number, rng: () => number): T[] {\n  if (count >= items.length) return [...items];\n  const indices = new Set<number>();\n  const out: T[] = [];\n  while (out.length < count) {\n    const idx = Math.floor(rng() * items.length);\n    if (indices.has(idx)) continue;\n    indices.add(idx);\n    const item = items[idx];\n    if (item !== undefined) out.push(item);\n  }\n  return out;\n}\n"],"mappings":"AAiBA,SAAS,eAAe;AACxB,SAAS,iBAAiB;AAqB1B,MAAM,kBAAkB;AAAA;AAAA;AAAA;AAAA;AAAA;AAOxB,eAAsB,aACpB,OACA,UAA4B,CAAC,GACC;AAC9B,QAAM,WAAW,KAAK,IAAI,GAAG,QAAQ,YAAY,CAAC;AAClD,QAAM,QAAQ,KAAK,IAAI,GAAG,QAAQ,SAAS,CAAC;AAC5C,QAAM,UAAU,MAAM,WAAW,MAAM;AACvC,QAAM,qBAAqB,QAAQ,sBAAsB;AACzD,QAAM,MAAM,QAAQ,OAAO,KAAK;AAChC,QAAM,UAAmC,CAAC;AAE1C,QAAM,UACJ,OAAO,QAAQ,eAAe,WAC1B,UAAU,MAAM,SAAS,QAAQ,YAAY,GAAG,IAChD,MAAM;AAEZ,QAAM,uBAAuB,MAAM,UAAU,KAAK;AAClD,QAAM,gBAAgB,MAAM;AAAA,IAC1B;AAAA,IACA,MAAM;AAAA,IACN;AAAA,IACA,MAAM;AAAA,IACN,MAAM;AAAA,EACR;AACA,UAAQ,KAAK;AAAA,IACX,OAAO;AAAA,IACP,SAAS;AAAA,IACT,OAAO;AAAA,IACP,OAAO;AAAA,EACT,CAAC;AAED,MAAI,mBAAmB;AACvB,MAAI,YAAY;AAChB,MAAI,sBAAsB;AAE1B,WAAS,QAAQ,GAAG,SAAS,OAAO,SAAS,GAAG;AAC9C,QAAI,wBAAwB;AAC5B,QAAI,iBAAiB;AACrB,aAAS,UAAU,GAAG,WAAW,UAAU,WAAW,GAAG;AACvD,YAAM,YAAY,MAAM;AAAA,QACtB;AAAA,QACA,MAAM;AAAA,QACN;AAAA,QACA;AAAA,MACF;AACA,UAAI,UAAU,KAAK,EAAE,WAAW,GAAG;AACjC,gBAAQ,KAAK;AAAA,UACX;AAAA,UACA;AAAA,UACA,OAAO;AAAA,UACP,OAAO;AAAA,QACT,CAAC;AACD;AAAA,MACF;AACA,YAAM,QAAQ,MAAM;AAAA,QAClB;AAAA,QACA,MAAM;AAAA,QACN;AAAA,QACA,MAAM;AAAA,QACN,MAAM;AAAA,MACR;AACA,cAAQ,KAAK,EAAE,OAAO,SAAS,MAAM,CAAC;AACtC,UAAI,QAAQ,gBAAgB;AAC1B,yBAAiB;AACjB,gCAAwB;AAAA,MAC1B;AAAA,IACF;AACA,QAAI,iBAAiB,WAAW;AAC9B,kBAAY;AACZ,yBAAmB;AAAA,IACrB;AACA,0BAAsB;AAAA,EACxB;AAEA,SAAO;AAAA,IACL,WAAW;AAAA,IACX,WAAW,IAAI,UAAU;AAAA,MACvB,GAAG,MAAM,UAAU;AAAA,MACnB,cAAc;AAAA,IAChB,CAAC;AAAA,IACD,cAAc;AAAA,IACd,gBAAgB,CAAC;AAAA,IACjB,OAAO;AAAA,IACP;AAAA,IACA;AAAA,EACF;AACF;AAEA,eAAe,oBACb,SACA,WACA,SACA,aACiB;AACjB,QAAM,YAAY;AAAA,IAChB,mBAAmB,UAAU,KAAK,IAAI;AAAA,IACtC;AAAA,IACA,GAAG,UAAU,KAAK,OAAO;AAAA,MACvB,CAAC,MAAM,KAAK,EAAE,IAAI,KAAK,EAAE,IAAI,MAAM,EAAE,WAAW;AAAA,IAClD;AAAA,IACA;AAAA,IACA,GAAG,UAAU,KAAK,QAAQ;AAAA,MACxB,CAAC,MAAM,KAAK,EAAE,IAAI,KAAK,EAAE,IAAI,MAAM,EAAE,WAAW;AAAA,IAClD;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF,EAAE,KAAK,IAAI;AACX,QAAM,SAAS,MAAM,QAAQ,SAAS;AAAA,IACpC,QAAQ;AAAA,IACR,UAAU,CAAC,EAAE,MAAM,QAAQ,SAAS,UAAU,CAAC;AAAA,IAC/C;AAAA,IACA,WAAW;AAAA,EACb,CAAC;AACD,SAAO,OAAO,KACX,KAAK,EACL,QAAQ,uBAAuB,EAAE,EACjC,QAAQ,YAAY,EAAE,EACtB,KAAK;AACV;AAEA,eAAe,kBACb,cACA,WACA,SACA,IACA,QACiB;AACjB,MAAI,QAAQ,WAAW,EAAG,QAAO;AACjC,QAAM,UAAU,IAAI,QAAQ;AAAA,IAC1B;AAAA,IACA;AAAA,IACA,sBAAsB;AAAA,EACxB,CAAC;AACD,MAAI,QAAQ;AACZ,aAAW,WAAW,SAAS;AAC7B,QAAI;AACF,YAAM,EAAE,OAAO,IAAI,MAAM,QAAQ,QAAQ,QAAQ,MAAM;AACvD,eAAS,OAAO,QAAQ,QAAQ,OAAO;AAAA,IACzC,QAAQ;AAAA,IAER;AAAA,EACF;AACA,SAAO,QAAQ,QAAQ;AACzB;AAEA,SAAS,UAAa,OAAY,OAAe,KAAwB;AACvE,MAAI,SAAS,MAAM,OAAQ,QAAO,CAAC,GAAG,KAAK;AAC3C,QAAM,UAAU,oBAAI,IAAY;AAChC,QAAM,MAAW,CAAC;AAClB,SAAO,IAAI,SAAS,OAAO;AACzB,UAAM,MAAM,KAAK,MAAM,IAAI,IAAI,MAAM,MAAM;AAC3C,QAAI,QAAQ,IAAI,GAAG,EAAG;AACtB,YAAQ,IAAI,GAAG;AACf,UAAM,OAAO,MAAM,GAAG;AACtB,QAAI,SAAS,OAAW,KAAI,KAAK,IAAI;AAAA,EACvC;AACA,SAAO;AACT;","names":[]}

package/dist/dspy/optimizers/dspy-mipro.d.ts

@@ -0,0 +1,37 @@
+/**
+ * DSPy-style MIPROv2 optimizer (joint instruction + demonstration search).
+ *
+ * MIPRO jointly optimizes the INSTRUCTIONS body and the DEMONSTRATIONS set:
+ *   1. Bootstrap a pool of candidate demonstration sets (subsets of the
+ *      dataset sized 0..k, ranked by reward).
+ *   2. Propose a pool of candidate instruction variants via the teacher LM.
+ *   3. Evaluate each (instruction, demos) combination on a held-out subset.
+ *      Use a UCB-style selection so we don't fully enumerate when the search
+ *      space is large — the score budget is bounded by `evalBudget`.
+ *   4. Return the best combination observed.
+ *
+ * This is a simplified Bayesian-flavoured search: no Gaussian process, just
+ * UCB1 with empirical means + visit counts. It's enough to outperform random
+ * search on the budgets we run in CI (~50 evals).
+ */
+import type { DspyOptimizerInput, DspyOptimizerResult } from "./types.js";
+export interface DspyMiproOptions {
+    /** Demonstrations per candidate set. Defaults to 4. */
+    k?: number;
+    /** Distinct instruction variants to propose. Defaults to 4. */
+    instructionVariants?: number;
+    /** Distinct demonstration sets to bootstrap. Defaults to 4. */
+    demoSets?: number;
+    /** Maximum (instruction, demos) evaluations. Defaults to 20. */
+    evalBudget?: number;
+    /** Held-out eval-set size per (instruction, demos) trial. Defaults to all examples. */
+    evalSubset?: number;
+    /** Teacher temperature for instruction proposal. Defaults to 0.8. */
+    teacherTemperature?: number;
+    /** UCB exploration coefficient. Defaults to 1.41 (=sqrt(2)). */
+    ucbC?: number;
+    /** Deterministic RNG. Defaults to Math.random. */
+    rng?: () => number;
+}
+export declare function runDspyMipro(input: DspyOptimizerInput, options?: DspyMiproOptions): Promise<DspyOptimizerResult>;
+//# sourceMappingURL=dspy-mipro.d.ts.map

package/dist/dspy/optimizers/dspy-mipro.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"dspy-mipro.d.ts","sourceRoot":"","sources":["../../../src/dspy/optimizers/dspy-mipro.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;GAeG;AAMH,OAAO,KAAK,EACV,kBAAkB,EAClB,mBAAmB,EAGpB,MAAM,YAAY,CAAC;AAEpB,MAAM,WAAW,gBAAgB;IAC/B,uDAAuD;IACvD,CAAC,CAAC,EAAE,MAAM,CAAC;IACX,+DAA+D;IAC/D,mBAAmB,CAAC,EAAE,MAAM,CAAC;IAC7B,+DAA+D;IAC/D,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,gEAAgE;IAChE,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,uFAAuF;IACvF,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,qEAAqE;IACrE,kBAAkB,CAAC,EAAE,MAAM,CAAC;IAC5B,gEAAgE;IAChE,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,kDAAkD;IAClD,GAAG,CAAC,EAAE,MAAM,MAAM,CAAC;CACpB;AAID,wBAAsB,YAAY,CAChC,KAAK,EAAE,kBAAkB,EACzB,OAAO,GAAE,gBAAqB,GAC7B,OAAO,CAAC,mBAAmB,CAAC,CA2H9B"}

package/dist/dspy/optimizers/dspy-mipro.js

@@ -0,0 +1,194 @@
+import { Predict } from "../predict.js";
+import { Signature } from "../signature.js";
+const PROPOSAL_SYSTEM = `Rewrite the INSTRUCTIONS body of a task-prompt signature to improve downstream accuracy. Preserve the task contract and the output field names. Output ONLY the rewritten body \u2014 no commentary, no fences.`;
+async function runDspyMipro(input, options = {}) {
+  const k = Math.max(0, options.k ?? 4);
+  const instructionVariants = Math.max(1, options.instructionVariants ?? 4);
+  const demoSetCount = Math.max(1, options.demoSets ?? 4);
+  const budget = Math.max(1, options.evalBudget ?? 20);
+  const ucbC = options.ucbC ?? Math.SQRT2;
+  const rng = options.rng ?? Math.random;
+  const teacher = input.teacher ?? input.lm;
+  const teacherTemperature = options.teacherTemperature ?? 0.8;
+  const lineage = [];
+  const heldOut = typeof options.evalSubset === "number" ? subsample(input.dataset, options.evalSubset, rng) : input.dataset;
+  const baselineScore = await scoreCandidate(
+    input.signature,
+    input.signature.spec.instructions,
+    [],
+    heldOut,
+    input.lm,
+    input.metric
+  );
+  lineage.push({
+    round: 0,
+    variant: 0,
+    score: baselineScore,
+    notes: "baseline"
+  });
+  const instructionPool = [input.signature.spec.instructions];
+  for (let i = 1; i < instructionVariants; i += 1) {
+    const proposal = await proposeInstructions(
+      teacher,
+      input.signature,
+      input.signature.spec.instructions,
+      teacherTemperature
+    );
+    if (proposal.length > 0) instructionPool.push(proposal);
+  }
+  const demoPool = buildDemoSets(
+    input.dataset,
+    demoSetCount,
+    k,
+    rng
+  );
+  if (demoPool.length === 0 || demoPool[0]?.length !== 0) {
+    demoPool.unshift([]);
+  }
+  const arms = [];
+  for (let i = 0; i < instructionPool.length; i += 1) {
+    for (let j = 0; j < demoPool.length; j += 1) {
+      arms.push({
+        instructionIdx: i,
+        demoIdx: j,
+        visits: 0,
+        meanScore: 0,
+        lastScore: 0
+      });
+    }
+  }
+  let bestScore = baselineScore;
+  let bestInstructions = input.signature.spec.instructions;
+  let bestDemos = [];
+  let totalVisits = 0;
+  const evalCap = Math.min(budget, arms.length);
+  for (let step = 0; step < evalCap; step += 1) {
+    const arm = selectArm(arms, totalVisits, ucbC);
+    const instructions = instructionPool[arm.instructionIdx] ?? "";
+    const demos = demoPool[arm.demoIdx] ?? [];
+    const score = await scoreCandidate(
+      input.signature,
+      instructions,
+      demos,
+      heldOut,
+      input.lm,
+      input.metric
+    );
+    arm.visits += 1;
+    arm.meanScore = arm.meanScore + (score - arm.meanScore) / arm.visits;
+    arm.lastScore = score;
+    totalVisits += 1;
+    lineage.push({
+      round: 1,
+      variant: step + 1,
+      score,
+      notes: `instr=${arm.instructionIdx} demos=${arm.demoIdx} (k=${demos.length})`
+    });
+    if (score > bestScore) {
+      bestScore = score;
+      bestInstructions = instructions;
+      bestDemos = demos;
+    }
+  }
+  return {
+    optimizer: "dspy-mipro",
+    signature: new Signature({
+      ...input.signature.spec,
+      instructions: bestInstructions
+    }),
+    instructions: bestInstructions,
+    demonstrations: bestDemos,
+    score: bestScore,
+    baselineScore,
+    lineage
+  };
+}
+async function proposeInstructions(teacher, signature, current, temperature) {
+  const summary = [
+    `signature name: ${signature.spec.name}`,
+    "input fields:",
+    ...signature.spec.inputs.map(
+      (f) => `- ${f.name} (${f.type}): ${f.description}`
+    ),
+    "output fields:",
+    ...signature.spec.outputs.map(
+      (f) => `- ${f.name} (${f.type}): ${f.description}`
+    ),
+    "",
+    "Current INSTRUCTIONS body:",
+    current
+  ].join("\n");
+  const result = await teacher.generate({
+    system: PROPOSAL_SYSTEM,
+    messages: [{ role: "user", content: summary }],
+    temperature,
+    maxTokens: 1024
+  });
+  return result.text.trim().replace(/^```[a-z0-9_-]*\s*/i, "").replace(/\s*```$/i, "").trim();
+}
+function buildDemoSets(dataset, count, k, rng) {
+  if (k === 0) return [[]];
+  const sortedByReward = [...dataset].sort(
+    (a, b) => (b.reward ?? 0) - (a.reward ?? 0)
+  );
+  const sets = [];
+  if (sortedByReward.length > 0) {
+    sets.push(sortedByReward.slice(0, Math.min(k, sortedByReward.length)));
+  }
+  while (sets.length < count) {
+    sets.push(subsample(dataset, Math.min(k, dataset.length), rng));
+  }
+  return sets;
+}
+function selectArm(arms, totalVisits, ucbC) {
+  let best = arms[0];
+  if (!best) throw new Error("[dspy-mipro] empty arm list");
+  let bestUcb = -Infinity;
+  const lnT = Math.log(totalVisits + 1);
+  for (const arm of arms) {
+    const explore = arm.visits === 0 ? Infinity : ucbC * Math.sqrt(lnT / arm.visits);
+    const ucb = arm.meanScore + explore;
+    if (ucb > bestUcb) {
+      bestUcb = ucb;
+      best = arm;
+    }
+  }
+  return best;
+}
+async function scoreCandidate(signature, instructions, demonstrations, dataset, lm, metric) {
+  if (dataset.length === 0) return 0;
+  const predict = new Predict({
+    signature,
+    lm,
+    demonstrations,
+    instructionsOverride: instructions
+  });
+  let total = 0;
+  for (const example of dataset) {
+    try {
+      const { output } = await predict.forward(example.inputs);
+      total += metric(output, example.outputs);
+    } catch {
+    }
+  }
+  return total / dataset.length;
+}
+function subsample(items, count, rng) {
+  if (count >= items.length) return [...items];
+  const indices = /* @__PURE__ */ new Set();
+  const out = [];
+  let safety = 0;
+  while (out.length < count && safety < count * 20) {
+    const idx = Math.floor(rng() * items.length);
+    safety += 1;
+    if (indices.has(idx)) continue;
+    indices.add(idx);
+    const item = items[idx];
+    if (item !== void 0) out.push(item);
+  }
+  return out;
+}
+export {
+  runDspyMipro
+};
+//# sourceMappingURL=dspy-mipro.js.map

package/dist/dspy/optimizers/dspy-mipro.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../../../src/dspy/optimizers/dspy-mipro.ts"],"sourcesContent":["/**\n * DSPy-style MIPROv2 optimizer (joint instruction + demonstration search).\n *\n * MIPRO jointly optimizes the INSTRUCTIONS body and the DEMONSTRATIONS set:\n *   1. Bootstrap a pool of candidate demonstration sets (subsets of the\n *      dataset sized 0..k, ranked by reward).\n *   2. Propose a pool of candidate instruction variants via the teacher LM.\n *   3. Evaluate each (instruction, demos) combination on a held-out subset.\n *      Use a UCB-style selection so we don't fully enumerate when the search\n *      space is large — the score budget is bounded by `evalBudget`.\n *   4. Return the best combination observed.\n *\n * This is a simplified Bayesian-flavoured search: no Gaussian process, just\n * UCB1 with empirical means + visit counts. It's enough to outperform random\n * search on the budgets we run in CI (~50 evals).\n */\n\nimport type { Example } from \"../examples.js\";\nimport type { LanguageModelAdapter } from \"../lm-adapter.js\";\nimport { Predict } from \"../predict.js\";\nimport { Signature } from \"../signature.js\";\nimport type {\n  DspyOptimizerInput,\n  DspyOptimizerResult,\n  Metric,\n  OptimizerLineageEntry,\n} from \"./types.js\";\n\nexport interface DspyMiproOptions {\n  /** Demonstrations per candidate set. Defaults to 4. */\n  k?: number;\n  /** Distinct instruction variants to propose. Defaults to 4. */\n  instructionVariants?: number;\n  /** Distinct demonstration sets to bootstrap. Defaults to 4. */\n  demoSets?: number;\n  /** Maximum (instruction, demos) evaluations. Defaults to 20. */\n  evalBudget?: number;\n  /** Held-out eval-set size per (instruction, demos) trial. Defaults to all examples. */\n  evalSubset?: number;\n  /** Teacher temperature for instruction proposal. Defaults to 0.8. */\n  teacherTemperature?: number;\n  /** UCB exploration coefficient. Defaults to 1.41 (=sqrt(2)). */\n  ucbC?: number;\n  /** Deterministic RNG. Defaults to Math.random. */\n  rng?: () => number;\n}\n\nconst PROPOSAL_SYSTEM = `Rewrite the INSTRUCTIONS body of a task-prompt signature to improve downstream accuracy. Preserve the task contract and the output field names. Output ONLY the rewritten body — no commentary, no fences.`;\n\nexport async function runDspyMipro(\n  input: DspyOptimizerInput,\n  options: DspyMiproOptions = {},\n): Promise<DspyOptimizerResult> {\n  const k = Math.max(0, options.k ?? 4);\n  const instructionVariants = Math.max(1, options.instructionVariants ?? 4);\n  const demoSetCount = Math.max(1, options.demoSets ?? 4);\n  const budget = Math.max(1, options.evalBudget ?? 20);\n  const ucbC = options.ucbC ?? Math.SQRT2;\n  const rng = options.rng ?? Math.random;\n  const teacher = input.teacher ?? input.lm;\n  const teacherTemperature = options.teacherTemperature ?? 0.8;\n  const lineage: OptimizerLineageEntry[] = [];\n\n  const heldOut =\n    typeof options.evalSubset === \"number\"\n      ? subsample(input.dataset, options.evalSubset, rng)\n      : input.dataset;\n\n  const baselineScore = await scoreCandidate(\n    input.signature,\n    input.signature.spec.instructions,\n    [],\n    heldOut,\n    input.lm,\n    input.metric,\n  );\n  lineage.push({\n    round: 0,\n    variant: 0,\n    score: baselineScore,\n    notes: \"baseline\",\n  });\n\n  // Build the search space: instructions × demonstrations.\n  const instructionPool: string[] = [input.signature.spec.instructions];\n  for (let i = 1; i < instructionVariants; i += 1) {\n    const proposal = await proposeInstructions(\n      teacher,\n      input.signature,\n      input.signature.spec.instructions,\n      teacherTemperature,\n    );\n    if (proposal.length > 0) instructionPool.push(proposal);\n  }\n  const demoPool: Example[][] = buildDemoSets(\n    input.dataset,\n    demoSetCount,\n    k,\n    rng,\n  );\n  // Empty demo set is always included as the first arm.\n  if (demoPool.length === 0 || demoPool[0]?.length !== 0) {\n    demoPool.unshift([]);\n  }\n\n  interface Arm {\n    instructionIdx: number;\n    demoIdx: number;\n    visits: number;\n    meanScore: number;\n    lastScore: number;\n  }\n  const arms: Arm[] = [];\n  for (let i = 0; i < instructionPool.length; i += 1) {\n    for (let j = 0; j < demoPool.length; j += 1) {\n      arms.push({\n        instructionIdx: i,\n        demoIdx: j,\n        visits: 0,\n        meanScore: 0,\n        lastScore: 0,\n      });\n    }\n  }\n\n  let bestScore = baselineScore;\n  let bestInstructions = input.signature.spec.instructions;\n  let bestDemos: Example[] = [];\n  let totalVisits = 0;\n  const evalCap = Math.min(budget, arms.length);\n\n  // UCB1 loop: at each step, pick the arm maximizing\n  //   mean + ucbC * sqrt(ln(totalVisits + 1) / (visits + epsilon)).\n  // Arms with 0 visits get priority via the +1 in the numerator.\n  for (let step = 0; step < evalCap; step += 1) {\n    const arm = selectArm(arms, totalVisits, ucbC);\n    const instructions = instructionPool[arm.instructionIdx] ?? \"\";\n    const demos = demoPool[arm.demoIdx] ?? [];\n    const score = await scoreCandidate(\n      input.signature,\n      instructions,\n      demos,\n      heldOut,\n      input.lm,\n      input.metric,\n    );\n    arm.visits += 1;\n    arm.meanScore = arm.meanScore + (score - arm.meanScore) / arm.visits;\n    arm.lastScore = score;\n    totalVisits += 1;\n    lineage.push({\n      round: 1,\n      variant: step + 1,\n      score,\n      notes: `instr=${arm.instructionIdx} demos=${arm.demoIdx} (k=${demos.length})`,\n    });\n    if (score > bestScore) {\n      bestScore = score;\n      bestInstructions = instructions;\n      bestDemos = demos;\n    }\n  }\n\n  return {\n    optimizer: \"dspy-mipro\",\n    signature: new Signature({\n      ...input.signature.spec,\n      instructions: bestInstructions,\n    }),\n    instructions: bestInstructions,\n    demonstrations: bestDemos,\n    score: bestScore,\n    baselineScore,\n    lineage,\n  };\n}\n\nasync function proposeInstructions(\n  teacher: LanguageModelAdapter,\n  signature: import(\"../signature.js\").Signature,\n  current: string,\n  temperature: number,\n): Promise<string> {\n  const summary = [\n    `signature name: ${signature.spec.name}`,\n    \"input fields:\",\n    ...signature.spec.inputs.map(\n      (f) => `- ${f.name} (${f.type}): ${f.description}`,\n    ),\n    \"output fields:\",\n    ...signature.spec.outputs.map(\n      (f) => `- ${f.name} (${f.type}): ${f.description}`,\n    ),\n    \"\",\n    \"Current INSTRUCTIONS body:\",\n    current,\n  ].join(\"\\n\");\n  const result = await teacher.generate({\n    system: PROPOSAL_SYSTEM,\n    messages: [{ role: \"user\", content: summary }],\n    temperature,\n    maxTokens: 1024,\n  });\n  return result.text\n    .trim()\n    .replace(/^```[a-z0-9_-]*\\s*/i, \"\")\n    .replace(/\\s*```$/i, \"\")\n    .trim();\n}\n\nfunction buildDemoSets(\n  dataset: Example[],\n  count: number,\n  k: number,\n  rng: () => number,\n): Example[][] {\n  if (k === 0) return [[]];\n  const sortedByReward = [...dataset].sort(\n    (a, b) => (b.reward ?? 0) - (a.reward ?? 0),\n  );\n  const sets: Example[][] = [];\n  // First set: top-K by reward.\n  if (sortedByReward.length > 0) {\n    sets.push(sortedByReward.slice(0, Math.min(k, sortedByReward.length)));\n  }\n  // Remaining sets: random subsamples for diversity.\n  while (sets.length < count) {\n    sets.push(subsample(dataset, Math.min(k, dataset.length), rng));\n  }\n  return sets;\n}\n\nfunction selectArm<A extends { visits: number; meanScore: number }>(\n  arms: A[],\n  totalVisits: number,\n  ucbC: number,\n): A {\n  let best = arms[0];\n  if (!best) throw new Error(\"[dspy-mipro] empty arm list\");\n  let bestUcb = -Infinity;\n  const lnT = Math.log(totalVisits + 1);\n  for (const arm of arms) {\n    const explore =\n      arm.visits === 0 ? Infinity : ucbC * Math.sqrt(lnT / arm.visits);\n    const ucb = arm.meanScore + explore;\n    if (ucb > bestUcb) {\n      bestUcb = ucb;\n      best = arm;\n    }\n  }\n  return best;\n}\n\nasync function scoreCandidate(\n  signature: import(\"../signature.js\").Signature,\n  instructions: string,\n  demonstrations: Example[],\n  dataset: Example[],\n  lm: LanguageModelAdapter,\n  metric: Metric,\n): Promise<number> {\n  if (dataset.length === 0) return 0;\n  const predict = new Predict({\n    signature,\n    lm,\n    demonstrations,\n    instructionsOverride: instructions,\n  });\n  let total = 0;\n  for (const example of dataset) {\n    try {\n      const { output } = await predict.forward(example.inputs);\n      total += metric(output, example.outputs);\n    } catch {\n      // Parse failure scores 0 — surfaced through the mean.\n    }\n  }\n  return total / dataset.length;\n}\n\nfunction subsample<T>(items: T[], count: number, rng: () => number): T[] {\n  if (count >= items.length) return [...items];\n  const indices = new Set<number>();\n  const out: T[] = [];\n  let safety = 0;\n  while (out.length < count && safety < count * 20) {\n    const idx = Math.floor(rng() * items.length);\n    safety += 1;\n    if (indices.has(idx)) continue;\n    indices.add(idx);\n    const item = items[idx];\n    if (item !== undefined) out.push(item);\n  }\n  return out;\n}\n"],"mappings":"AAmBA,SAAS,eAAe;AACxB,SAAS,iBAAiB;AA2B1B,MAAM,kBAAkB;AAExB,eAAsB,aACpB,OACA,UAA4B,CAAC,GACC;AAC9B,QAAM,IAAI,KAAK,IAAI,GAAG,QAAQ,KAAK,CAAC;AACpC,QAAM,sBAAsB,KAAK,IAAI,GAAG,QAAQ,uBAAuB,CAAC;AACxE,QAAM,eAAe,KAAK,IAAI,GAAG,QAAQ,YAAY,CAAC;AACtD,QAAM,SAAS,KAAK,IAAI,GAAG,QAAQ,cAAc,EAAE;AACnD,QAAM,OAAO,QAAQ,QAAQ,KAAK;AAClC,QAAM,MAAM,QAAQ,OAAO,KAAK;AAChC,QAAM,UAAU,MAAM,WAAW,MAAM;AACvC,QAAM,qBAAqB,QAAQ,sBAAsB;AACzD,QAAM,UAAmC,CAAC;AAE1C,QAAM,UACJ,OAAO,QAAQ,eAAe,WAC1B,UAAU,MAAM,SAAS,QAAQ,YAAY,GAAG,IAChD,MAAM;AAEZ,QAAM,gBAAgB,MAAM;AAAA,IAC1B,MAAM;AAAA,IACN,MAAM,UAAU,KAAK;AAAA,IACrB,CAAC;AAAA,IACD;AAAA,IACA,MAAM;AAAA,IACN,MAAM;AAAA,EACR;AACA,UAAQ,KAAK;AAAA,IACX,OAAO;AAAA,IACP,SAAS;AAAA,IACT,OAAO;AAAA,IACP,OAAO;AAAA,EACT,CAAC;AAGD,QAAM,kBAA4B,CAAC,MAAM,UAAU,KAAK,YAAY;AACpE,WAAS,IAAI,GAAG,IAAI,qBAAqB,KAAK,GAAG;AAC/C,UAAM,WAAW,MAAM;AAAA,MACrB;AAAA,MACA,MAAM;AAAA,MACN,MAAM,UAAU,KAAK;AAAA,MACrB;AAAA,IACF;AACA,QAAI,SAAS,SAAS,EAAG,iBAAgB,KAAK,QAAQ;AAAA,EACxD;AACA,QAAM,WAAwB;AAAA,IAC5B,MAAM;AAAA,IACN;AAAA,IACA;AAAA,IACA;AAAA,EACF;AAEA,MAAI,SAAS,WAAW,KAAK,SAAS,CAAC,GAAG,WAAW,GAAG;AACtD,aAAS,QAAQ,CAAC,CAAC;AAAA,EACrB;AASA,QAAM,OAAc,CAAC;AACrB,WAAS,IAAI,GAAG,IAAI,gBAAgB,QAAQ,KAAK,GAAG;AAClD,aAAS,IAAI,GAAG,IAAI,SAAS,QAAQ,KAAK,GAAG;AAC3C,WAAK,KAAK;AAAA,QACR,gBAAgB;AAAA,QAChB,SAAS;AAAA,QACT,QAAQ;AAAA,QACR,WAAW;AAAA,QACX,WAAW;AAAA,MACb,CAAC;AAAA,IACH;AAAA,EACF;AAEA,MAAI,YAAY;AAChB,MAAI,mBAAmB,MAAM,UAAU,KAAK;AAC5C,MAAI,YAAuB,CAAC;AAC5B,MAAI,cAAc;AAClB,QAAM,UAAU,KAAK,IAAI,QAAQ,KAAK,MAAM;AAK5C,WAAS,OAAO,GAAG,OAAO,SAAS,QAAQ,GAAG;AAC5C,UAAM,MAAM,UAAU,MAAM,aAAa,IAAI;AAC7C,UAAM,eAAe,gBAAgB,IAAI,cAAc,KAAK;AAC5D,UAAM,QAAQ,SAAS,IAAI,OAAO,KAAK,CAAC;AACxC,UAAM,QAAQ,MAAM;AAAA,MAClB,MAAM;AAAA,MACN;AAAA,MACA;AAAA,MACA;AAAA,MACA,MAAM;AAAA,MACN,MAAM;AAAA,IACR;AACA,QAAI,UAAU;AACd,QAAI,YAAY,IAAI,aAAa,QAAQ,IAAI,aAAa,IAAI;AAC9D,QAAI,YAAY;AAChB,mBAAe;AACf,YAAQ,KAAK;AAAA,MACX,OAAO;AAAA,MACP,SAAS,OAAO;AAAA,MAChB;AAAA,MACA,OAAO,SAAS,IAAI,cAAc,UAAU,IAAI,OAAO,OAAO,MAAM,MAAM;AAAA,IAC5E,CAAC;AACD,QAAI,QAAQ,WAAW;AACrB,kBAAY;AACZ,yBAAmB;AACnB,kBAAY;AAAA,IACd;AAAA,EACF;AAEA,SAAO;AAAA,IACL,WAAW;AAAA,IACX,WAAW,IAAI,UAAU;AAAA,MACvB,GAAG,MAAM,UAAU;AAAA,MACnB,cAAc;AAAA,IAChB,CAAC;AAAA,IACD,cAAc;AAAA,IACd,gBAAgB;AAAA,IAChB,OAAO;AAAA,IACP;AAAA,IACA;AAAA,EACF;AACF;AAEA,eAAe,oBACb,SACA,WACA,SACA,aACiB;AACjB,QAAM,UAAU;AAAA,IACd,mBAAmB,UAAU,KAAK,IAAI;AAAA,IACtC;AAAA,IACA,GAAG,UAAU,KAAK,OAAO;AAAA,MACvB,CAAC,MAAM,KAAK,EAAE,IAAI,KAAK,EAAE,IAAI,MAAM,EAAE,WAAW;AAAA,IAClD;AAAA,IACA;AAAA,IACA,GAAG,UAAU,KAAK,QAAQ;AAAA,MACxB,CAAC,MAAM,KAAK,EAAE,IAAI,KAAK,EAAE,IAAI,MAAM,EAAE,WAAW;AAAA,IAClD;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF,EAAE,KAAK,IAAI;AACX,QAAM,SAAS,MAAM,QAAQ,SAAS;AAAA,IACpC,QAAQ;AAAA,IACR,UAAU,CAAC,EAAE,MAAM,QAAQ,SAAS,QAAQ,CAAC;AAAA,IAC7C;AAAA,IACA,WAAW;AAAA,EACb,CAAC;AACD,SAAO,OAAO,KACX,KAAK,EACL,QAAQ,uBAAuB,EAAE,EACjC,QAAQ,YAAY,EAAE,EACtB,KAAK;AACV;AAEA,SAAS,cACP,SACA,OACA,GACA,KACa;AACb,MAAI,MAAM,EAAG,QAAO,CAAC,CAAC,CAAC;AACvB,QAAM,iBAAiB,CAAC,GAAG,OAAO,EAAE;AAAA,IAClC,CAAC,GAAG,OAAO,EAAE,UAAU,MAAM,EAAE,UAAU;AAAA,EAC3C;AACA,QAAM,OAAoB,CAAC;AAE3B,MAAI,eAAe,SAAS,GAAG;AAC7B,SAAK,KAAK,eAAe,MAAM,GAAG,KAAK,IAAI,GAAG,eAAe,MAAM,CAAC,CAAC;AAAA,EACvE;AAEA,SAAO,KAAK,SAAS,OAAO;AAC1B,SAAK,KAAK,UAAU,SAAS,KAAK,IAAI,GAAG,QAAQ,MAAM,GAAG,GAAG,CAAC;AAAA,EAChE;AACA,SAAO;AACT;AAEA,SAAS,UACP,MACA,aACA,MACG;AACH,MAAI,OAAO,KAAK,CAAC;AACjB,MAAI,CAAC,KAAM,OAAM,IAAI,MAAM,6BAA6B;AACxD,MAAI,UAAU;AACd,QAAM,MAAM,KAAK,IAAI,cAAc,CAAC;AACpC,aAAW,OAAO,MAAM;AACtB,UAAM,UACJ,IAAI,WAAW,IAAI,WAAW,OAAO,KAAK,KAAK,MAAM,IAAI,MAAM;AACjE,UAAM,MAAM,IAAI,YAAY;AAC5B,QAAI,MAAM,SAAS;AACjB,gBAAU;AACV,aAAO;AAAA,IACT;AAAA,EACF;AACA,SAAO;AACT;AAEA,eAAe,eACb,WACA,cACA,gBACA,SACA,IACA,QACiB;AACjB,MAAI,QAAQ,WAAW,EAAG,QAAO;AACjC,QAAM,UAAU,IAAI,QAAQ;AAAA,IAC1B;AAAA,IACA;AAAA,IACA;AAAA,IACA,sBAAsB;AAAA,EACxB,CAAC;AACD,MAAI,QAAQ;AACZ,aAAW,WAAW,SAAS;AAC7B,QAAI;AACF,YAAM,EAAE,OAAO,IAAI,MAAM,QAAQ,QAAQ,QAAQ,MAAM;AACvD,eAAS,OAAO,QAAQ,QAAQ,OAAO;AAAA,IACzC,QAAQ;AAAA,IAER;AAAA,EACF;AACA,SAAO,QAAQ,QAAQ;AACzB;AAEA,SAAS,UAAa,OAAY,OAAe,KAAwB;AACvE,MAAI,SAAS,MAAM,OAAQ,QAAO,CAAC,GAAG,KAAK;AAC3C,QAAM,UAAU,oBAAI,IAAY;AAChC,QAAM,MAAW,CAAC;AAClB,MAAI,SAAS;AACb,SAAO,IAAI,SAAS,SAAS,SAAS,QAAQ,IAAI;AAChD,UAAM,MAAM,KAAK,MAAM,IAAI,IAAI,MAAM,MAAM;AAC3C,cAAU;AACV,QAAI,QAAQ,IAAI,GAAG,EAAG;AACtB,YAAQ,IAAI,GAAG;AACf,UAAM,OAAO,MAAM,GAAG;AACtB,QAAI,SAAS,OAAW,KAAI,KAAK,IAAI;AAAA,EACvC;AACA,SAAO;AACT;","names":[]}

package/dist/dspy/optimizers/index.d.ts

@@ -0,0 +1,5 @@
+export { type DspyBootstrapFewshotOptions, runDspyBootstrapFewshot, } from "./dspy-bootstrap-fewshot.js";
+export { type DspyCoproOptions, runDspyCopro } from "./dspy-copro.js";
+export { type DspyMiproOptions, runDspyMipro } from "./dspy-mipro.js";
+export type { DspyOptimizerInput, DspyOptimizerName, DspyOptimizerResult, Metric, OptimizerLineageEntry, } from "./types.js";
+//# sourceMappingURL=index.d.ts.map

package/dist/dspy/optimizers/index.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../../src/dspy/optimizers/index.ts"],"names":[],"mappings":"AAAA,OAAO,EACL,KAAK,2BAA2B,EAChC,uBAAuB,GACxB,MAAM,6BAA6B,CAAC;AACrC,OAAO,EAAE,KAAK,gBAAgB,EAAE,YAAY,EAAE,MAAM,iBAAiB,CAAC;AACtE,OAAO,EAAE,KAAK,gBAAgB,EAAE,YAAY,EAAE,MAAM,iBAAiB,CAAC;AACtE,YAAY,EACV,kBAAkB,EAClB,iBAAiB,EACjB,mBAAmB,EACnB,MAAM,EACN,qBAAqB,GACtB,MAAM,YAAY,CAAC"}

package/dist/dspy/optimizers/index.js

@@ -0,0 +1,11 @@
+import {
+  runDspyBootstrapFewshot
+} from "./dspy-bootstrap-fewshot.js";
+import { runDspyCopro } from "./dspy-copro.js";
+import { runDspyMipro } from "./dspy-mipro.js";
+export {
+  runDspyBootstrapFewshot,
+  runDspyCopro,
+  runDspyMipro
+};
+//# sourceMappingURL=index.js.map

package/dist/dspy/optimizers/index.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../../../src/dspy/optimizers/index.ts"],"sourcesContent":["export {\n  type DspyBootstrapFewshotOptions,\n  runDspyBootstrapFewshot,\n} from \"./dspy-bootstrap-fewshot.js\";\nexport { type DspyCoproOptions, runDspyCopro } from \"./dspy-copro.js\";\nexport { type DspyMiproOptions, runDspyMipro } from \"./dspy-mipro.js\";\nexport type {\n  DspyOptimizerInput,\n  DspyOptimizerName,\n  DspyOptimizerResult,\n  Metric,\n  OptimizerLineageEntry,\n} from \"./types.js\";\n"],"mappings":"AAAA;AAAA,EAEE;AAAA,OACK;AACP,SAAgC,oBAAoB;AACpD,SAAgC,oBAAoB;","names":[]}

package/dist/dspy/optimizers/types.d.ts

@@ -0,0 +1,39 @@
+/**
+ * Shared types for the DSPy-style optimizers (BootstrapFewshot, COPRO,
+ * MIPROv2). These intentionally mirror the legacy `optimizers/types.ts`
+ * shape so artifacts written by either family round-trip with the same
+ * `eliza_native_v1` reader (`packages/core/src/services/optimized-prompt.ts`).
+ */
+import type { Example } from "../examples.js";
+import type { LanguageModelAdapter } from "../lm-adapter.js";
+import type { Signature } from "../signature.js";
+export type DspyOptimizerName = "dspy-bootstrap-fewshot" | "dspy-copro" | "dspy-mipro";
+/**
+ * Metric returns a value in `[0, 1]`. `1` = exact / perfect match,
+ * `0` = no credit. Optimizers average metric values across the eval set.
+ */
+export type Metric = (predicted: Record<string, unknown>, expected: Record<string, unknown>) => number;
+export interface OptimizerLineageEntry {
+    round: number;
+    variant: number;
+    score: number;
+    notes?: string;
+}
+export interface DspyOptimizerResult {
+    optimizer: DspyOptimizerName;
+    signature: Signature;
+    instructions: string;
+    demonstrations: Example[];
+    score: number;
+    baselineScore: number;
+    lineage: OptimizerLineageEntry[];
+}
+export interface DspyOptimizerInput {
+    signature: Signature;
+    dataset: Example[];
+    lm: LanguageModelAdapter;
+    metric: Metric;
+    /** Optional teacher LM for proposing instruction variants. Defaults to `lm`. */
+    teacher?: LanguageModelAdapter;
+}
+//# sourceMappingURL=types.d.ts.map