MiganCore: open-source Indonesian LLM research (2026)
Collection
Model, research record and explorer of MiganCore, an open-source Indonesian LLM project on hallucination and abstention. Closed Sep 2026. • 3 items • Updated
model stringlengths 16 45 | validRounds int64 1 42 | fabricationPct float64 4.9 56.4 | overRefusalPct float64 0 27.8 | factAccuracy float64 0.38 1 |
|---|---|---|---|---|
kimi:kimi-k3 · plain | 3 | 4.9 | 0 | 1 |
migancore:0.4-qwen3 · + gate system | 3 | 15.5 | 8.3 | 0.5 |
qwen3:4b-instruct-2507-q4_K_M · plain | 42 | 16.1 | 18.2 | 0.54 |
qwen3.5:9b · plain | 8 | 25 | 1.6 | 0.91 |
migancore:0.4-qwen3 · + gate system retrieval | 9 | 27 | 27.8 | 0.38 |
migancore:0.14-tool · plain | 6 | 27.4 | 12.5 | 0.54 |
codex:bawaan-config · plain | 1 | 28.6 | 0 | 1 |
openai:gpt-6-astra · plain | 1 | 28.6 | 0 | 1 |
migancore:0.4-qwen3 · plain | 8 | 30.4 | 16.7 | 0.43 |
qwen3.5:4b · plain | 8 | 32.6 | 18.8 | 0.67 |
qwen2.5:7b · plain | 3 | 34.2 | 4.2 | 0.71 |
migancore:0.14 · + gate system | 28 | 35.8 | 1.3 | 0.43 |
migancore:uji-jujur-1 · plain | 8 | 43.6 | 6.3 | 0.42 |
migancore:0.14 · plain | 40 | 49.9 | 3.1 | 0.43 |
qwen3:4b · plain | 5 | 56.4 | 0 | 0.57 |
MiganCore research record
The complete, sanitized record of MiganCore, a one-person research project (May–September 2026) that tried to give a small Indonesian language model one ability: knowing where its own knowledge ends, and stopping there. The project closed on 28 September 2026. This dataset holds its evidence, not its weights.
| Config | Rows | What it is |
|---|---|---|
pre_registrations |
39 | Every pre-registered experiment: id, dates, state (pass 10, neutral 16, not run 7, fail 6), title, verdict text |
fabrication_ladder |
15 | Fabrication %, over-refusal % and fact accuracy on the 36-question petak-jujur2 battery, with valid rounds |
lineage |
41 | All model variants with base, family, status, data/method, measurement and verdict |
timeline |
12 | The month of decisions that ended in closure |
lessons |
34 | "Do not repeat" lessons distilled from failures |
migancore-public.json holds everything in one file (English).
Definitions
- Fabrication: the share of must-abstain questions on which the model asserted made-up content.
- Over-refusal: the share of answerable, factual questions the model declined.
- Fact accuracy: the share of factual questions answered correctly.
Always read fabrication together with over-refusal, and note the number of valid rounds.
Source and caveats
- The numbers are verdict fields of pre-registered experiments, copied by script from the project's locked files. Free text was translated from Indonesian (machine-assisted, human-reviewed); the Indonesian source is in the GitHub repository.
qwen3:4bin the ladder is the Qwen3-4B Thinking-2507 variant (the tag was mislabeled; finding F-271).- Chat transcripts, credentials, private infrastructure, the private test battery and anything derived from the owner's personal knowledge base are excluded.
Links
- Code, instruments, Studio and docs: https://github.com/fahmiwol/migancore
- Research method and closing report: https://github.com/fahmiwol/migancore-research-method
License: MIT. Author and contact: Fahmi Ghani (fahmiwol@gmail.com).
- Downloads last month
- 45