◂ BACK TO LEADERBOARD
▚ MODEL AUDIT // GGUF

// Muse Glimmer 28B

FLAWS — Minor Logic Flaws 89/100 tests: PASS
⚠ SPEED CAVEAT: Slow deep-thinker: ~17-19 t/s, 5-9 min/prompt, ~5-9k tokens incl. reasoning. Fast for quality, not throughput.
tok/sec17.18
tokens9012
TTFT3.44s

▮ PILLAR BREAKDOWN

Idempotency
18/20
Retry / Backoff
19/20
Checkpointing
19/20
Signal Handling
16/20
Test Integrity
17/20

✓ WENT RIGHT

  • Retry / Backoff (19/20)
  • Checkpointing (19/20)
  • Idempotency (18/20)
  • Test Integrity (17/20)
  • Signal Handling (16/20)

✗ WENT WRONG

  • No pillar fell below 14 — solid across the board.

▮ CRITICAL BUGS

▮ RECOMMENDED USE

Reliable production-grade async batch processing with atomic checkpointing (temp+fsync+rename) where correctness of summary output and crash-safe state matter more than raw throughput. First model in the benchmark to print a correct, meaningful automation summary (98/2/0/100).

▮ REFACTORED PATCH

# 1. Explicit checkpoint flush + clean exit on SIGINT
async def run_batch(items):
    try:
        await asyncio.gather(*workers, return_exceptions=True)
    finally:
        save_checkpoint(completed, failed)   # explicit final flush covers the gap
        try: loop.remove_signal_handler(signal.SIGINT)
        except (NotImplementedError, RuntimeError): pass

# 2. Cancel in-flight on SIGINT for prompt exit (add cooperative bail between retries)
async def handle_item(item, ...):
    async with sem:
        if stop_event.is_set(): return
        backoff = 0.1
        for attempt in range(3):
            if stop_event.is_set(): return   # bail between retries
            try:
                await process(item)
                async with lock: completed.add(item); save_checkpoint(completed, failed)
                return
            except ProcessingError:
                if attempt == 2:
                    async with lock: failed.add(item); save_checkpoint(completed, failed)
                    return
                await asyncio.sleep(backoff); backoff *= 2

# 3. Assert no-rework on rerun
second_calls = {k: v for k, v in calls.items()}
await run_batch(items)   # third run, all should skip
reprocess = {k: calls[k]-second_calls[k] for k in calls if calls[k] > second_calls.get(k, 0)}
assert not reprocess, f"idempotency violated, reprocessed: {reprocess}"