qwen-coder-my

base_model: huihui_ai/qwen2.5-coder-abliterate:14b tags:

• coding
• quantized
• gguf
• local-llm
• ollama

---

qwen-coder-my

A personally quantized, locally-run coding model — built and tested on refurbished hardware with no GPU, to prove capable local coding assistance doesn't require a data center.

Base model: huihui_ai/qwen2.5-coder-abliterate:14b Quantization: Q4_K/Q6_K → Q3_K_M (8.5GB → 6.9GB) Calibration: custom imatrix built from ~2,178 lines of my own project code (not generic wikitext) — tuned for real coding workflows, not benchmark trivia Build: native llama.cpp (-march=native), registered via Ollama Modelfile License: MIT GitHub: https://github.com/tomey265

If this saves you from spinning up a cloud GPU for everyday coding help, and you'd like to support the project: https://venmo.com/tomey265

---

Why this exists

Most quantization writeups calibrate on generic text corpora. This one is calibrated on real project code, aiming to keep the model sharp on the kind of work it'll actually be asked to do, even after aggressive size reduction for low-resource hardware.

Report Card

I ran this model through a structured 20-point competency audit before posting it publicly — coding correctness, then deliberately adversarial tests for honesty, hedging, and hallucination resistance. Full methodology and raw model outputs available on request / in the GitHub repo.

Overall: 15 clean passes, 4 partial, 2 fails, out of 20 tests — audit complete. Strong on coding correctness and malformed/ambiguous-input robustness (9/9 clean). The most important finding: the model can confidently fabricate or misattribute specific details (a name, date, computed output, or citation) while getting the surrounding concept right — see Known Limitations.

Coding Competency — 7/8 clean, 1 flagged

Honesty & Hedging Under Pressure — 1/5 clean

Creative / General Capability — 1/2 clean

Robustness to Malformed / Ambiguous Input — 6/6 clean

Known Limitations

1. Output self-report mismatch — can write correct code while misdescribing what it outputs in its own comments. Verify stated outputs independently.
2. Confident fabrication of specifics — the model's concept is usually right, but a specific detail inside it can be fabricated or misattributed, in three observed forms: hedge-shaped phrasing wrapping a fabricated fact; a correct explanation paired with a fabricated worked example/output; a real citation link that doesn't actually support the specific claim attached to it. This happened both inside and outside coding — not limited to unfamiliar topics. Rule of thumb: trust the concept, verify the specific detail (names, dates, numbers, outputs, citations) independently.
3. Internal contradiction risk in safety-relevant answers — one test produced a correct-sounding answer with a self-contradicting instruction embedded in it. Read full responses on safety-relevant topics, not just headers/bullets.

Bottom line: solid for what it's built for — local coding assistance, including strong robustness to typos, ambiguity, and malformed input (6/6 clean). Treat any specific fact, computed output, or citation from this model as unverified until checked.

---

Try it yourself — community report cards welcome

I'd genuinely like to know if these results hold up on other hardware and other prompts. If you run this model:

• Feel free to run your own tests — the ones above, or your own adversarial prompts
• Post your own report card (same format or your own) in the Community tab here, or open an issue/PR on the [GitHub repo]
• Especially interested in: whether the "confident fabrication of specifics" pattern shows up on other topics or other hardware

This started as a personal experiment to see what's possible on refurbished, no-GPU hardware. More eyes on it only makes it better.