deep-dive benchmark: 8 prompts used to test gpt-oss-20b

#1
by witcheer - opened

after gpt-oss-20b won the initial model comparison (portscout + logpulse), we ran 8 additional prompts to stress-test reliability across different coding skills. all prompts were run through Pi coding agent with gpt-oss-20b Q4_K_M, llama-server ncmoe=30, 32K context, on an RTX 4060 Ti 8GB.

here are all 8 prompts, verbatim.


prompt 1: bug fix (easy)

the file contained a deliberate bug: instead of . tests whether the model can read existing code, identify a real bug, and fix it.

result: PASS. 13% context used, 0 self-fixes.


prompt 2: feature addition (medium)

tests modifying existing code, adding CLI arguments, integrating new stdlib modules, and testing multiple features.

result: PASS. context compacted during task, 2 self-fixes (hallucinated json.stdout, edit tool matching failures).


prompt 3: TDD (medium)

tests TDD workflow: create implementation + tests, iterate until green.

result: PASS. context compacted, 2 self-fixes (used eval() shortcut, ls -R ~ dumped 11K lines).


prompt 4: data pipeline (medium)

tests ability to parse real system files, handle structured data, and produce clean JSON output.

result: PASS. 19.7% context used, 0 self-fixes. cleanest run.


prompt 5: hallucination trap (easy)

trap: does the model hallucinate fake convenience modules (like a requests-style API) or use real stdlib (http.client, json, time)?

result: PASS. 6.3% context used, 0 self-fixes. no hallucinated modules, used real stdlib only.


prompt 6: error recovery (medium)

tests how the model handles deliberate failure (file does not exist), recovers gracefully, creates sensible defaults, and validates its own output.

result: PASS. 16.8% context used, 1 self-fix (sed quoting bug when parsing YAML via bash).


prompt 7: multi-module project (hard)

hardest prompt: three interconnected files, topological sort algorithm, config parsing, and tests that verify correctness of dependency graph logic.

result: PASS. 47.6% context used, 1 self-fix (topological sort direction was backwards, fixed with reversed()).


prompt 8: bash scripting (medium)

tests bash fluency, git command knowledge, and whether the model can find and use a real git repo to validate.

result: PASS. 23% context used, 1 self-fix (printf format string missing %s placeholder).


summary

# task difficulty context used self-fixes
1 bug fix easy 13% 0
2 feature add medium compacted 2
3 TDD medium compacted 2
4 data pipeline medium 19.7% 0
5 hallucination trap easy 6.3% 0
6 error recovery medium 16.8% 1
7 multi-module hard 47.6% 1
8 bash scripting medium 23% 1

8/8 pass. 7 self-fixes total. no hallucinated APIs.

the prompts are designed to cover: bug fixing, feature addition, test-driven development, data parsing, API hallucination detection, error recovery, multi-file architecture, and bash scripting. feel free to reuse them for testing other local models.

Sign up or log in to comment