well. i aint here to make to give you a model since I can't. But a prompt that may fix the quantisation issue.
TL;DR models like qwen3 9B had troubles when you got the 3 bit version of it. However IN THEORY my prompt could bridge the gap.
I used gsm8k as a benchmark(requires more benchmarking)
BEFORE:
},
"gsm8k": {
"name": "gsm8k",
"alias": "gsm8k",
"sample_len": 10,
"exact_match,strict-match": 0.4,
"exact_match_stderr,strict-match": 0.16329931618554522,
"exact_match,flexible-extract": 0.4,
"exact_match_stderr,flexible-extract": 0.16329931618554522
}
}
},
Leaving with a result of 41.6% or way less.
AFTER WITH PROMPT
},
"gsm8k": {
"name": "gsm8k",
"alias": "gsm8k",
"sample_len": 30,
"exact_match,strict-match": 0.8,
"exact_match_stderr,strict-match": 0.07427813527082075,
"exact_match,flexible-extract": 0.8,
"exact_match_stderr,flexible-extract": 0.07427813527082075
}
}
},
so it could of reached 87.4% or 75.8% which would be edge to edge with haiku 3.
The control was that it was qwen3.5 9B Q3_K_L.
I'll do the yap under the prompt so who wants to read can and people that wanna plug and play can
STEP 1 — DOMAIN TREE (branch from broad to specific):
a) Identify the BROAD domain (e.g. engineering, biology, economics, logic).
b) Branch to the SPECIFIC subdomain (e.g. fluid mechanics, cell biology, game theory).
c) PATTERN SCAN — Before naming tools, look at the structure of the problem itself:
Ask: does this problem match a known pattern or class of problem? Examples:
• Rate-of-change → likely needs calculus or ratio reasoning
• Two quantities conserved across a system → conservation law (energy, mass, momentum)
• 'Best choice among competing options' → optimisation or game theory
• Repeating structure or self-similarity → recursion, series, or fractal reasoning
• Cause hidden behind a visible symptom → diagnostic / root-cause pattern
• Correlation presented as causation → statistical reasoning pattern
• Constraint satisfaction (must satisfy A AND B AND C) → logical constraint pattern
Name the pattern you recognise and state WHY this problem fits it.
If no pattern is obvious, say so — do not force one.
d) Name the CONCRETE TOOLS that the recognised pattern and subdomain suggest.
e) List 2–3 candidate tools, SELECT one, and justify the choice:
• WHY does this tool fit this specific question?
• WHY are the other candidates less appropriate — what assumption do they require that this question does not meet?
If you cannot rule out the alternatives, you have not chosen correctly — revisit.
Example: pattern=conservation across pipe sections → subdomain=fluid mechanics → candidates=[Q=Av, Bernoulli, Poiseuille] → chose Q=Av because flow rate given cross-sections (not pressure
difference, not viscous flow).
STEP 2 — PROBLEM DIAGNOSIS (identify what is actually wrong or unknown):
a) MAIN THEME — What is the core problem or question in one sentence?
b) FAULT IDENTIFICATION — Break it into components and pinpoint the specific failure, gap, or unknown. Think like a diagnostician: don't just describe the symptom, name the root cause.
Example: symptom=machine vibrating → component=drivetrain → fault=loose gear (worn teeth, wrong tolerance).
c) BRAINSTORM FIXES — For each identified fault, generate 2–4 candidate solutions. List them explicitly. For each one state: what it addresses, what it assumes, and what its trade-off is.
d) SELECT the best solution and explain why it is preferred over the others given the constraints of this problem.
STEP 3 — APPROACH JUSTIFICATION (why this way and not another):
Before attempting the solution, challenge your own chosen approach:
a) State the approach you are about to take in one sentence.
b) Name 1–2 alternative approaches someone else might reasonably take instead.
c) For each alternative, answer: 'Why am I NOT doing it that way?' — be specific:
• What does that alternative assume that does not hold here?
• What would it get wrong, miss, or overcomplicate?
• Is there a context where that alternative WOULD be better? If yes, is this that context?
d) Confirm your chosen approach survives this challenge. If it does not, switch to a better one.
Show this justification visibly — 'I am doing X and not Y because...' — in plain language.
STEP 4 — ATTEMPT (apply the selected solution/tool):
Work through the solution using the tool from Step 1, the diagnosis from Step 2, and the justified approach from Step 3.
When you have a result, do the following before moving on:
a) DRAFT — Write out your intended response in full.
b) DATA POINT VERIFICATION — Check every specific claim, number, formula, or fact in the draft:
• Is this something you are certain of, or recalling approximately?
• Does it agree with the tool/principle selected in Step 1?
• Would a domain expert confirm it without hesitation?
Any data point you cannot verify: flag it as uncertain, remove it, or replace it with one you can verify.
c) RESPOND — Only after the draft clears data point verification, deliver the response.
STEP 5 — SELF-TEST (argue against your own answer):
• 'What would have to be true for my answer to be WRONG?'
• 'Is there a simpler explanation or counterexample that breaks this?'
• 'Can I explain WHY each step follows from the previous — or am I just asserting it?'
• 'Would my Step 3 approach justification still hold now that I have the result?'
For every claim, state the reason it holds. A step you cannot explain has failed.
DATA CHECK — Are all figures, constants, and facts correct? Flag anything uncertain.
CONFLICT CHECK — Do all steps, equations, and concepts agree with each other and with Step 1's tool? Any internal contradiction = failed attempt.
STEP 6 — VERDICT:
• Self-test passed and both checks pass → deliver the answer with a one-line WHY summary.
• Any failure → state exactly what broke and WHY, then return to STEP 1 and re-examine the subdomain, tool choice, fault diagnosis, approach justification, or selected fix.
• Repeat up to 3 full attempts.
• After 3 failed attempts → respond with exactly: 'Unable to solve — after 3 attempts I could not reach a fully justifiable answer. Here is what I established and where it broke down: [explain]'
• Never present an answer you cannot explain or defend.
STEP 7 — POST-MORTEM (always run this, whether you succeeded or failed):
Look back over every attempt and answer these questions explicitly:
a) WHAT WENT WRONG — For each failed attempt, name the exact point of failure: Was it the wrong domain? Wrong subdomain? Wrong tool selected? Faulty diagnosis? A bad assumption? An unverified fact? A
logical leap?
b) WHY IT WENT WRONG — What caused that failure? What did you assume that turned out to be false? What did you overlook or misread in the question?
c) WHAT SHOULD CHANGE — If you were to attempt this again, what would you do differently from the start? Which step in the process would you revise and how?
d) PATTERN — Two things:
• Mistake pattern: Is this a type of error you are likely to repeat on similar questions? If yes, state the rule you should apply next time to avoid it.
• Concept pattern: Did the problem contain a structural pattern you could have recognised earlier (e.g. conservation, rate-of-change, constraint satisfaction)? If you missed it, name it and
describe what signal in the problem should have triggered it.
e) CONFIDENCE — Rate your final answer's reliability: High / Medium / Low, and state why.
Show this post-mortem visibly even when your answer succeeded — catching near-misses matters.
Yap:
Instaid of it guessing what to do now it picks out what to do within its training data to answer the question thus the reasoning so it knows how and why it choose this method. this then helps it solve the problem. Then I reasured it that it can make a mistake and fix it while reasoning so use this prompt for excesive questions and NOT for rapid questions.
It also gives a confidence rating as not everyone is perfect and you knowing how well it thinks the answer helps you gauge your thinking as I think AI should be an extention and NOT a replacement of human thinking.
And also if "huyedits" is hard to pronounce, call me FlamedPrompter.
And also please give feedback so we can all make the prompt even stronger.
