// LLM NER annotator. Runs inside the harness JS eval kernel because `completion()` only // exists there. Invoke from an eval cell: // // globalThis.ANNOTATE = { // jobs: [{ requests: "/…/shard-000.jsonl", responses: "/…/shard-000.jsonl" }, …], // cache: "/annotation/work/cache", // model: "default", concurrency: 16, // }; // await (0, eval)(await Bun.file("/scripts/annotation/annotate.js").text()); // // A single `{requests, responses}` pair may be given at the top level instead of `jobs`. // Every batch of every job goes into ONE work queue, so `concurrency` is the real number of // in-flight model calls no matter how the work is split across files; shard boundaries cost // nothing in throughput. // // Indirect eval runs in global scope, so `completion` resolves. Everything lives inside one // async IIFE so the file can be evaluated repeatedly without lexical redeclaration errors. // // Per request line: cache hit -> reuse; else call the model, verify the returned id set // equals the requested one, retry once on mismatch, then split the batch in half and // re-issue each half. A call that throws (rate limit, transport) is retried with backoff and // then recorded as failed; failures are never cached, so re-running resumes exactly the // batches that did not land. (async () => { const cfg = globalThis.ANNOTATE; if (!cfg) throw new Error("set globalThis.ANNOTATE before evaluating annotate.js"); if (typeof completion === "undefined") { throw new Error("completion() is not visible; paste this file's body into the eval cell"); } const model = cfg.model ?? "default"; const concurrency = cfg.concurrency ?? 8; const attempts = cfg.attempts ?? 3; const jobs = cfg.jobs ?? [{ requests: cfg.requests, responses: cfg.responses }]; const INSTRUCTION = "Annotate every sentence. Return one item per id."; const { mkdir, writeFile } = await import("node:fs/promises"); const path = await import("node:path"); await mkdir(cfg.cache, { recursive: true }); const queue = []; for (const [j, job] of jobs.entries()) { const text = await Bun.file(job.requests).text(); const reqs = text.split("\n").filter((l) => l.trim()).map((l) => JSON.parse(l)); job.results = new Array(reqs.length); await mkdir(path.dirname(job.responses), { recursive: true }); reqs.forEach((req, i) => queue.push({ job: j, index: i, req })); } const stats = { cached: 0, called: 0, retried: 0, errored: 0, failed: 0 }; const sleep = (ms) => new Promise((r) => setTimeout(r, ms)); const parse = (raw) => (typeof raw === "string" ? JSON.parse(raw) : raw); const bodyFor = (req, ids) => { const byId = new Map(req.sentences.map((s) => [s.id, s.text])); return INSTRUCTION + "\n\n" + ids.map((id) => `${id}: ${byId.get(id)}`).join("\n"); }; // One model call, retried through transport/rate-limit errors before giving up. const ask = async (req, ids) => { for (let a = 1; ; a++) { try { stats.called += 1; const out = parse( await completion(bodyFor(req, ids), { model, system: req.system, schema: req.schema }).wait(), ); const items = Array.isArray(out?.items) ? out.items : []; const want = new Set(ids); const got = new Set(items.map((i) => i.id)); const ok = got.size === want.size && [...want].every((id) => got.has(id)); return { items: items.filter((i) => want.has(i.id)), ok }; } catch (err) { stats.errored += 1; if (a >= attempts) { console.log(` call failed on ${req.batch} after ${a} attempts: ${err?.message ?? err}`); return { items: [], ok: false, threw: true }; } await sleep(1000 * 2 ** (a - 1)); } } }; const run = async (req, ids, depth = 0) => { let attempt = await ask(req, ids); if (attempt.ok) return { items: attempt.items, error: null }; // A thrown call (transport, rate limit, content refusal) is not necessarily the whole // batch's fault: a single sentence can trigger a refusal. Bisect so the rest still lands. if (attempt.threw && ids.length === 1) return { items: [], error: "call-failed" }; if (!attempt.threw && depth === 0) { stats.retried += 1; attempt = await ask(req, ids); if (attempt.ok) return { items: attempt.items, error: null }; } if (ids.length > 1 && (attempt.threw || depth < 2)) { const mid = Math.ceil(ids.length / 2); const [ra, rb] = [ await run(req, ids.slice(0, mid), depth + 1), await run(req, ids.slice(mid), depth + 1), ]; const errs = [ra.error, rb.error].filter(Boolean); return { items: [...ra.items, ...rb.items], error: errs.length ? errs[0] : null }; } console.log(` id-mismatch on ${req.batch} ids=${ids.join(",")}`); return { items: attempt.items, error: "id-mismatch" }; }; let next = 0; const worker = async () => { for (;;) { const slot = next++; if (slot >= queue.length) return; const { job: j, index, req } = queue[slot]; const cacheFile = path.join(cfg.cache, `${req.cache_key}.json`); if (await Bun.file(cacheFile).exists()) { jobs[j].results[index] = JSON.parse(await Bun.file(cacheFile).text()); stats.cached += 1; continue; } if (cfg.cacheOnly) { jobs[j].results[index] = { batch: req.batch, model, prompt_hash: req.prompt_hash, items: [], error: "uncached", failed_ids: req.ids, }; stats.failed += 1; continue; } const { items, error } = await run(req, req.ids); const record = { batch: req.batch, model, prompt_hash: req.prompt_hash, items, error, ...(error ? { failed_ids: req.ids.filter((id) => !items.some((it) => it.id === id)) } : {}), }; jobs[j].results[index] = record; if (error) stats.failed += 1; else await writeFile(cacheFile, JSON.stringify(record), "utf8"); } }; const t0 = Date.now(); await Promise.all(Array.from({ length: concurrency }, worker)); for (const job of jobs) { await writeFile(job.responses, job.results.map((r) => JSON.stringify(r)).join("\n") + "\n", "utf8"); } const secs = (Date.now() - t0) / 1000; const line = `jobs=${jobs.length} batches=${queue.length} cached=${stats.cached} called=${stats.called}` + ` retried=${stats.retried} call_errors=${stats.errored} failed=${stats.failed}` + ` concurrency=${concurrency} elapsed=${secs.toFixed(1)}s (${(queue.length / secs).toFixed(2)} batches/s)`; console.log(line); return line; })()