rlhf-feedback-loop 0.5.0

package/scripts/dpo-optimizer.js

@@ -0,0 +1,208 @@
+'use strict';
+/**
+ * DPO Batch Optimizer (DPO-02)
+ *
+ * Builds (chosen, rejected) preference pairs from feedback-log.jsonl memories,
+ * computes DPO log-ratio adjustments using Thompson Sampling posteriors,
+ * and writes dpo-model.json to RLHF_FEEDBACK_DIR.
+ *
+ * Does NOT call any external API. Pure offline batch optimization.
+ *
+ * Exports: run, buildPreferencePairs, applyDpoAdjustments, dpoLogRatio
+ */
+
+const fs = require('fs');
+const path = require('path');
+const os = require('os');
+
+const DPO_BETA = 0.1;
+
+// ---------------------------------------------------------------------------
+// dpoLogRatio — DPO preference adjustment formula
+// ---------------------------------------------------------------------------
+
+/**
+ * Compute DPO adjustment in range [-1, +1] from chosen and rejected weights.
+ *
+ * @param {number} chosenWeight  - Time-decay weight for chosen (positive) outcome
+ * @param {number} rejectedWeight - Time-decay weight for rejected (negative) outcome
+ * @param {number} [beta=0.1]   - Temperature parameter (lower = more aggressive)
+ * @returns {number} Adjustment in [-1, +1]
+ */
+function dpoLogRatio(chosenWeight, rejectedWeight, beta) {
+  const b = (beta !== undefined && beta !== null) ? beta : DPO_BETA;
+  const cw = Math.max(chosenWeight, 0.01);
+  const rw = Math.max(rejectedWeight, 0.01);
+  const logRatio = Math.log(cw) - Math.log(rw);
+  const sigmoid = 1.0 / (1.0 + Math.exp(-b * logRatio));
+  return (sigmoid - 0.5) * 2;
+}
+
+// ---------------------------------------------------------------------------
+// buildPreferencePairs — groups DPO pairs by category from feedbackDir
+// ---------------------------------------------------------------------------
+
+/**
+ * Build preference pairs grouped by category.
+ *
+ * Uses buildDpoPairs() from export-dpo-pairs.js (do NOT reimplement).
+ * Reads memory-log.jsonl from feedbackDir to get error + learning memories.
+ *
+ * @param {string} feedbackDir - Directory containing memory-log.jsonl
+ * @returns {Object} Map of category → [{ chosen, rejected }]
+ */
+function buildPreferencePairs(feedbackDir) {
+  const { buildDpoPairs, readJSONL } = require('./export-dpo-pairs');
+  const memoryLogPath = path.join(feedbackDir, 'memory-log.jsonl');
+  const memories = readJSONL(memoryLogPath);
+
+  const errors = memories.filter((m) => m.category === 'error');
+  const learnings = memories.filter((m) => m.category === 'learning');
+
+  const result = buildDpoPairs(errors, learnings);
+
+  // Group pairs by category (inferred from matchedKeys or tags)
+  const grouped = {};
+  for (const pair of result.pairs) {
+    const keys = (pair.metadata && pair.metadata.matchedKeys) || [];
+    const category = keys.length > 0 ? keys[0] : 'uncategorized';
+    if (!grouped[category]) grouped[category] = [];
+    grouped[category].push({
+      chosen: pair.metadata
+        ? { id: pair.metadata.learningId, content: pair.chosen, timestamp: null }
+        : { content: pair.chosen },
+      rejected: pair.metadata
+        ? { id: pair.metadata.errorId, content: pair.rejected, timestamp: null }
+        : { content: pair.rejected },
+      metadata: pair.metadata,
+    });
+  }
+
+  return grouped;
+}
+
+// ---------------------------------------------------------------------------
+// applyDpoAdjustments — mutates Thompson model with DPO posterior adjustments
+// ---------------------------------------------------------------------------
+
+/**
+ * Apply DPO adjustments to Thompson Sampling posteriors and save the model.
+ *
+ * CRITICAL: calls ts.saveModel(model, modelPath) after all mutations.
+ *
+ * @param {string} modelPath - Path to feedback_model.json
+ * @param {Object} pairs     - Map of category → [{ chosen, rejected }] from buildPreferencePairs
+ * @returns {Object} adjustments - Map of category → { pairs, avg_adjustment }
+ */
+function applyDpoAdjustments(modelPath, pairs) {
+  const ts = require('./thompson-sampling');
+  const model = ts.loadModel(modelPath);
+  const adjustments = {};
+
+  for (const [cat, catPairs] of Object.entries(pairs)) {
+    if (!catPairs || catPairs.length === 0) continue;
+
+    // Ensure category exists in model
+    if (!model.categories[cat]) {
+      model.categories[cat] = { alpha: 1.0, beta: 1.0, samples: 0, last_updated: null };
+    }
+
+    let total = 0;
+    for (const pair of catPairs) {
+      const chosenTs = (pair.chosen && pair.chosen.timestamp) || null;
+      const rejectedTs = (pair.rejected && pair.rejected.timestamp) || null;
+      const cw = ts.timeDecayWeight(chosenTs);
+      const rw = ts.timeDecayWeight(rejectedTs);
+      total += dpoLogRatio(cw, rw);
+    }
+
+    const avg = total / catPairs.length;
+
+    if (avg > 0) {
+      model.categories[cat].alpha += avg * catPairs.length * 0.5;
+    } else {
+      model.categories[cat].beta += Math.abs(avg) * catPairs.length * 0.5;
+    }
+
+    adjustments[cat] = {
+      pairs: catPairs.length,
+      avg_adjustment: Math.round(avg * 10000) / 10000,
+    };
+  }
+
+  // CRITICAL: save after all mutations (Pitfall 2 from RESEARCH.md)
+  ts.saveModel(model, modelPath);
+
+  return adjustments;
+}
+
+// ---------------------------------------------------------------------------
+// run — top-level batch DPO optimization entry point
+// ---------------------------------------------------------------------------
+
+/**
+ * Run the full DPO optimization batch:
+ * 1. Build preference pairs from memory-log.jsonl
+ * 2. Apply DPO adjustments to Thompson model
+ * 3. Write dpo-model.json to feedbackDir
+ *
+ * @param {Object} [opts]
+ * @param {string} [opts.feedbackDir] - Override RLHF_FEEDBACK_DIR
+ * @param {string} [opts.modelPath]   - Override Thompson model path
+ * @returns {{ adjustments: Object, pairs_processed: number }}
+ */
+function run(opts) {
+  const options = opts || {};
+  const feedbackDir = options.feedbackDir ||
+    process.env.RLHF_FEEDBACK_DIR ||
+    path.join(os.homedir(), '.claude', 'memory', 'feedback');
+  const modelPath = options.modelPath ||
+    path.join(process.cwd(), '.claude', 'memory', 'feedback', 'feedback_model.json');
+
+  const pairs = buildPreferencePairs(feedbackDir);
+
+  const pairsProcessed = Object.values(pairs).reduce((sum, arr) => sum + arr.length, 0);
+
+  let adjustments = {};
+  if (pairsProcessed > 0) {
+    adjustments = applyDpoAdjustments(modelPath, pairs);
+  }
+
+  const dpoModel = {
+    generated: new Date().toISOString(),
+    pairs_processed: pairsProcessed,
+    adjustments,
+  };
+
+  const dpoModelPath = path.join(feedbackDir, 'dpo-model.json');
+  if (!fs.existsSync(feedbackDir)) {
+    fs.mkdirSync(feedbackDir, { recursive: true });
+  }
+  fs.writeFileSync(dpoModelPath, `${JSON.stringify(dpoModel, null, 2)}\n`);
+
+  console.log(`DPO optimization complete: ${pairsProcessed} pairs processed`);
+  if (Object.keys(adjustments).length > 0) {
+    console.log('Adjustments:', JSON.stringify(adjustments, null, 2));
+  } else {
+    console.log('No adjustment pairs found (empty or no overlapping memories)');
+  }
+
+  return { adjustments, pairs_processed: pairsProcessed };
+}
+
+// ---------------------------------------------------------------------------
+// CLI entry point
+// ---------------------------------------------------------------------------
+
+if (require.main === module && process.argv.includes('--run')) {
+  run().catch((e) => {
+    console.error(e);
+    process.exit(1);
+  });
+}
+
+// ---------------------------------------------------------------------------
+// Exports
+// ---------------------------------------------------------------------------
+
+module.exports = { run, buildPreferencePairs, applyDpoAdjustments, dpoLogRatio };

package/scripts/export-dpo-pairs.js

@@ -0,0 +1,316 @@
+#!/usr/bin/env node
+/**
+ * DPO Preference Pair Exporter
+ *
+ * Transforms error + learning memories into DPO JSONL triples.
+ */
+
+const fs = require('fs');
+const path = require('path');
+const { traceForDpoPair, aggregateTraces } = require('./code-reasoning');
+
+const PROJECT_ROOT = path.join(__dirname, '..');
+const FEEDBACK_DIR = process.env.RLHF_FEEDBACK_DIR || path.join(PROJECT_ROOT, '.claude', 'memory', 'feedback');
+const DEFAULT_LOCAL_MEMORY_LOG = path.join(FEEDBACK_DIR, 'memory-log.jsonl');
+
+function readJSONL(filePath) {
+  if (!fs.existsSync(filePath)) return [];
+  const raw = fs.readFileSync(filePath, 'utf-8').trim();
+  if (!raw) return [];
+  return raw
+    .split('\n')
+    .map((line) => {
+      try {
+        return JSON.parse(line);
+      } catch {
+        return null;
+      }
+    })
+    .filter(Boolean);
+}
+
+function extractDomainKeys(memory) {
+  const keys = new Set();
+  const genericTags = new Set(['feedback', 'positive', 'negative']);
+
+  if (Array.isArray(memory.tags)) {
+    for (const tag of memory.tags) {
+      if (!genericTags.has(tag)) keys.add(tag);
+    }
+  }
+
+  const titleWords = (memory.title || '')
+    .replace(/^(MISTAKE|SUCCESS|ERROR|LEARNING|PREFERENCE):\s*/i, '')
+    .toLowerCase()
+    .split(/\s+/)
+    .filter((w) => w.length > 3)
+    .slice(0, 3);
+
+  for (const word of titleWords) keys.add(word);
+
+  return Array.from(keys);
+}
+
+function domainOverlap(keysA, keysB) {
+  const setB = new Set(keysB);
+  let overlap = 0;
+  for (const key of keysA) {
+    if (setB.has(key)) overlap++;
+  }
+  return overlap;
+}
+
+function inferPrompt(error, learning) {
+  const shared = (error.tags || []).filter((t) => (learning.tags || []).includes(t));
+  if (shared.length > 0) {
+    return `Task domain: ${shared.join(', ')}. How should the agent handle this scenario?`;
+  }
+
+  const scenario = (error.title || '').replace(/^(MISTAKE|ERROR):\s*/i, '').trim();
+  if (scenario) return `Scenario: ${scenario}. What is the better response?`;
+  return 'How should the agent respond in this situation?';
+}
+
+function getRubricWeightedScore(memory) {
+  if (!memory || !memory.rubricSummary) return null;
+  const score = Number(memory.rubricSummary.weightedScore);
+  if (!Number.isFinite(score)) return null;
+  return score;
+}
+
+function buildRubricDelta(error, learning) {
+  const errorScore = getRubricWeightedScore(error);
+  const learningScore = getRubricWeightedScore(learning);
+  if (errorScore == null && learningScore == null) return null;
+  const delta = (learningScore != null && errorScore != null)
+    ? Math.round((learningScore - errorScore) * 1000) / 1000
+    : null;
+  return {
+    learningWeightedScore: learningScore,
+    errorWeightedScore: errorScore,
+    weightedDelta: delta,
+    errorFailingCriteria: error && error.rubricSummary ? error.rubricSummary.failingCriteria || [] : [],
+    learningFailingCriteria: learning && learning.rubricSummary ? learning.rubricSummary.failingCriteria || [] : [],
+  };
+}
+
+function buildDpoPairs(errors, learnings) {
+  const pairs = [];
+  const usedErrors = new Set();
+  const usedLearnings = new Set();
+
+  const errorKeys = errors.map((e) => ({ memory: e, keys: extractDomainKeys(e) }));
+  const learningKeys = learnings.map((l) => ({ memory: l, keys: extractDomainKeys(l) }));
+
+  for (const err of errorKeys) {
+    let best = null;
+    let bestScore = 0;
+    let bestOverlap = 0;
+
+    for (const learn of learningKeys) {
+      if (usedLearnings.has(learn.memory.id)) continue;
+      const overlap = domainOverlap(err.keys, learn.keys);
+      const rubric = buildRubricDelta(err.memory, learn.memory);
+      const rubricDelta = rubric && rubric.weightedDelta != null ? rubric.weightedDelta : 0;
+      const score = overlap + Math.max(0, rubricDelta);
+      if (score > bestScore) {
+        best = learn;
+        bestScore = score;
+        bestOverlap = overlap;
+      }
+    }
+
+    if (best && bestScore > 0 && bestOverlap > 0) {
+      pairs.push({
+        prompt: inferPrompt(err.memory, best.memory),
+        chosen: best.memory.content,
+        rejected: err.memory.content,
+        metadata: {
+          errorId: err.memory.id,
+          learningId: best.memory.id,
+          matchScore: bestScore,
+          overlapScore: domainOverlap(err.keys, best.keys),
+          matchedKeys: err.keys.filter((k) => best.keys.includes(k)),
+          errorTitle: err.memory.title,
+          learningTitle: best.memory.title,
+          rubric: buildRubricDelta(err.memory, best.memory),
+        },
+      });
+      usedErrors.add(err.memory.id);
+      usedLearnings.add(best.memory.id);
+    }
+  }
+
+  return {
+    pairs,
+    unpairedErrors: errors.filter((e) => !usedErrors.has(e.id)),
+    unpairedLearnings: learnings.filter((l) => !usedLearnings.has(l.id)),
+  };
+}
+
+function toJSONL(pairs) {
+  return `${pairs.map((p) => JSON.stringify(p)).join('\n')}\n`;
+}
+
+function exportDpoFromMemories(memories) {
+  const errors = memories.filter((m) => m.category === 'error');
+  const learnings = memories.filter((m) => m.category === 'learning');
+  const result = buildDpoPairs(errors, learnings);
+
+  const traces = result.pairs.map((pair) => traceForDpoPair(pair));
+  const reasoning = aggregateTraces(traces);
+
+  const pairsWithTraces = result.pairs.map((pair, i) => ({
+    ...pair,
+    metadata: {
+      ...pair.metadata,
+      reasoningTrace: {
+        traceId: traces[i].traceId,
+        confidence: traces[i].summary.confidence,
+        passed: traces[i].summary.passed,
+        verified: traces[i].summary.verified,
+        refuted: traces[i].summary.refuted,
+        edgeCases: traces[i].edgeCases,
+      },
+    },
+  }));
+
+  return {
+    pairs: pairsWithTraces,
+    unpairedErrors: result.unpairedErrors,
+    unpairedLearnings: result.unpairedLearnings,
+    errors,
+    learnings,
+    reasoning,
+    jsonl: toJSONL(pairsWithTraces),
+  };
+}
+
+function parseArgs(argv) {
+  const args = {};
+  argv.forEach((arg) => {
+    if (!arg.startsWith('--')) return;
+    const [key, ...rest] = arg.slice(2).split('=');
+    args[key] = rest.length ? rest.join('=') : true;
+  });
+  return args;
+}
+
+function runCli() {
+  const args = parseArgs(process.argv.slice(2));
+
+  if (args.test) {
+    runTests();
+    return;
+  }
+
+  let memories = [];
+
+  if (args.input) {
+    const raw = fs.readFileSync(args.input, 'utf-8');
+    const parsed = JSON.parse(raw);
+    memories = Array.isArray(parsed) ? parsed : parsed.memories || [];
+  } else if (args['from-local']) {
+    memories = readJSONL(DEFAULT_LOCAL_MEMORY_LOG);
+  } else {
+    console.error('Provide --input=<path-to-json> or --from-local');
+    process.exit(1);
+  }
+
+  const result = exportDpoFromMemories(memories);
+  const jsonl = result.jsonl;
+
+  if (args.output) {
+    fs.writeFileSync(args.output, jsonl);
+    console.error(`Wrote ${result.pairs.length} DPO pairs to ${args.output}`);
+  } else {
+    process.stdout.write(jsonl);
+  }
+
+  console.error(`Errors=${result.errors.length} Learnings=${result.learnings.length} Pairs=${result.pairs.length}`);
+  console.error(`Unpaired errors=${result.unpairedErrors.length} Unpaired learnings=${result.unpairedLearnings.length}`);
+}
+
+function runTests() {
+  let passed = 0;
+  let failed = 0;
+
+  function assert(condition, name) {
+    if (condition) {
+      passed++;
+      console.log(`  PASS ${name}`);
+    } else {
+      failed++;
+      console.log(`  FAIL ${name}`);
+    }
+  }
+
+  console.log('\nexport-dpo-pairs.js tests\n');
+
+  const errors = [
+    {
+      id: 1,
+      title: 'MISTAKE: Claimed done with no test proof',
+      content: 'Claimed completion without running tests.',
+      category: 'error',
+      tags: ['verification', 'feedback'],
+      rubricSummary: {
+        weightedScore: 0.32,
+        failingCriteria: ['verification_evidence'],
+        failingGuardrails: ['testsPassed'],
+      },
+    },
+    {
+      id: 2,
+      title: 'MISTAKE: Generic mismatch',
+      content: 'No matching learning memory for this domain.',
+      category: 'error',
+      tags: ['unique-tag'],
+    },
+  ];
+
+  const learnings = [
+    {
+      id: 10,
+      title: 'SUCCESS: Always run tests before completion claims',
+      content: 'Run tests and include output before saying complete.',
+      category: 'learning',
+      tags: ['verification', 'feedback'],
+      rubricSummary: {
+        weightedScore: 0.89,
+        failingCriteria: [],
+        failingGuardrails: [],
+      },
+    },
+  ];
+
+  const result = buildDpoPairs(errors, learnings);
+  assert(result.pairs.length === 1, 'one pair built from overlapping domain keys');
+  assert(result.unpairedErrors.length === 1, 'unpaired error left when no match exists');
+  assert(result.unpairedLearnings.length === 0, 'no unpaired learnings');
+
+  const jsonl = toJSONL(result.pairs);
+  assert(jsonl.endsWith('\n'), 'JSONL output ends with newline');
+
+  const parsed = JSON.parse(jsonl.trim());
+  assert(parsed.prompt.includes('verification'), 'inferred prompt includes shared domain');
+  assert(parsed.metadata.rubric.weightedDelta > 0, 'rubric delta metadata is attached');
+
+  console.log(`\nResults: ${passed} passed, ${failed} failed\n`);
+  process.exit(failed > 0 ? 1 : 0);
+}
+
+module.exports = {
+  readJSONL,
+  extractDomainKeys,
+  domainOverlap,
+  inferPrompt,
+  buildDpoPairs,
+  toJSONL,
+  exportDpoFromMemories,
+  DEFAULT_LOCAL_MEMORY_LOG,
+};
+
+if (require.main === module) {
+  runCli();
+}