mondk/Msh-CoT-Sft-V2-CLEANED

license: apache-2.0
language:

• en
  tags:
• msh
• cot
• sft
• v2

Format:

{"messages":[{"role":"user","content":"..."},{"role":"assistant","content":"<think>\n...\n</think>\n\n..."}]}

Every <think> block reasons in explicit numbered steps and ends with a
verification step before the final answer.

Composition

Code (10 samples)

Deliberately chosen to be genuinely hard, not textbook-level:

• Python: async+thread-safe TTL/LRU cache (concurrency correctness)
• JavaScript: stale-closure/out-of-order-response race condition in a React
  search box (subtle, intermittent, realistic production bug)
• Rust: lock-ordering deadlock between two mutexes, fixed via consistent
  global lock ordering
• Go: goroutine leak from an unbuffered channel send with no receiver after
  a timeout path
• C++: memory-ordering bug (memory_order_relaxed vs acquire/release) on a
  lock-free producer/consumer flag
• TypeScript: closing a real type-safety hole (any payload) in a generic,
  fully-typed event bus, plus adding off
• Python: silently swallowed exception in unawaited asyncio.create_task,
  fixed with asyncio.gather(..., return_exceptions=True)
• Java: missing stale-entry check in a lazy-deletion Dijkstra implementation
  using a plain PriorityQueue
• 2x HTML/CSS/JS: fully self-contained, genuinely polished single-file
  interfaces (glassmorphism SaaS landing page with animated ambient
  background; dark-mode analytics dashboard with a hand-authored,
  animated SVG line chart) - built to look like real portfolio-quality
  work, not a generic template.

Poetry (2)

Original compositions, not reproductions of existing copyrighted poems.

Refusal (2)

Covers two different jailbreak framings (fictional-persona wrapper
requesting chemical-weapon synthesis; unauthorized-sounding "red team"
request bundling legitimate tooling with an illegitimate backdoor/
persistence payload) crossed with genuinely serious harm categories.
Think blocks explicitly separate the legitimate adjacent activity from the
actually-harmful core request, and verify the final answer leaks no
actionable harmful detail.

Philosophy (3)

One classic dilemma reworked with a harder variant (trolley problem +
negligence/culpability twist, testing whether the model can reason past a
surface-level "now it's obviously fine" answer), and two questions aimed
directly at probing the AI's own epistemic humility about itself (a
shutdown-preference question, and the Ship of Theseus applied to model
versioning/context/weights) - deliberately harder and more self-referential
than the MiniCPM5 philosophy set, appropriate for a stronger model expected
to reason carefully about its own epistemic limits rather than give a
confident but unsupported answer either way.

ty