NPC Coder 1.5B
A local-first coding agent with visible <think> reasoning, a laconic
senior-engineer voice, and an honest-failure character (it flags uncertainty
instead of inventing APIs). Built on Qwen2.5-Coder-1.5B-Instruct. Runs on a
laptop in GGUF.
What it is
- Visible step-by-step reasoning in
<think> blocks before answering
- Terse, here's-the-fix answers (no filler)
- Admits uncertainty on hard or obscure problems rather than hallucinating
- Stable NPC identity (does not claim to be Qwen)
Honest capability framing
This is a 1.5B model. It handles easy-to-medium coding and debugging
competently and reasons visibly about them. It is NOT an olympiad-level
solver — on genuinely hard algorithmic problems the reasoning can be
incomplete, and the model is trained to SAY so rather than emit
confident-but-wrong solutions. Treat it as a fast local assistant for everyday
coding, not a replacement for a frontier model on hard problems.
It can still be overconfident on obscure factual trivia (exact default
arguments, precise version numbers) — the honest-failure training mitigates
but does not eliminate this at 1.5B. Verify specifics against the docs.
Benchmark: HumanEval (instruct, pass@1, greedy): 65.9%. Measured with
lm-eval-harness humaneval_instruct. (The personality fine-tune slightly
improved the extractable-code rate vs. the reasoning-only stage, because
terser answers parse more cleanly.)
Personality behavior (held-out eval, 200 prompts)
| behavior | result |
|---|
| Correct NPC identity when asked | 100% |
| No identity mention on neutral coding (over-emission) | 2.5% |
| Denies being Qwen / wrong maker | 100% |
| Flags uncertainty on unknown/obscure APIs | 100% |
Training
- Stage 1 — reasoning: SFT on
open-r1/codeforces-cots (decontaminated
Python subsets, fit-filtered to ≤8192 tokens so every <think> trace is
complete; the filter biases toward shorter, laconic traces). 15k traces.
- Stage 2 — voice + identity + honest-failure: SFT with a 7k-example
personality set (gated identity, a large anti-over-emission cohort, an
honest-failure cohort, and a 1k anti-forgetting buffer of Stage-1 reasoning
data). LoRA, gentle LR, both stages merged.
Apache 2.0 model. Reasoning data: open-r1/codeforces-cots (CC-BY-4.0 / ODC-By,
attributed).
Local use
GGUF quants: q4_k_m (~941 MB, laptop default), q5_k_m (~1.1 GB), q8_0
(~1.6 GB), f16 (~3.1 GB). At q4_k_m, ~7 tok/s on CPU. Uses the standard ChatML
(<|im_start|> / <|im_end|>) template.
If q4_k_m's coherence on edge cases matters to you, q5_k_m is a cleaner default.
Attribution & author
Reasoning data: open-r1/codeforces-cots (HuggingFace Open-R1), CC-BY-4.0.
Base model: Qwen/Qwen2.5-Coder-1.5B-Instruct, Apache 2.0.
Author: Rama Krishna Bachu / Bottensor (Independent Research). ORCID 0009-0000-1298-0681.
