File size: 1,835 Bytes
3f164ee fb9784f | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 | ---
language:
- tr
tags:
- devim
- evaluation
- benchmark-methodology
- turkish
- research
pretty_name: DEVİM Evaluation Research
---
# DEVİM Eval
This repository documents DEVİM's evaluation methodology and evaluator failures. It does **not** release hidden holdout items.
A central research lesson in DEVİM is that evaluator quality is part of the scientific problem: a model can improve language modeling while a capability instrument stays flat, and a seemingly useful QA task can turn out to be solvable by copying.
## Frozen V1.8 evaluator
Public-safe facts:
- Evaluator: **DEVIM_TR_CAPABILITY_V1_8**
- Forced-choice items: **2,000**
- Validator: **32 / 32 PASS**
- Hard contamination detected: **0**
- Blind audit: **75 CLEAN / 5 AWKWARD / 0 HARD**
The hidden item bank itself remains private to preserve evaluation integrity.
## Why evaluator failures are published
DEVİM's QA diagnostic exposed a copy confound: the interface could not cleanly distinguish target capability from mechanical extraction. That experiment was treated as evaluator evidence, not as a model victory.
The current public principle is:
**research metrics are not automatically leaderboard benchmarks.**
A trustworthy public benchmark requires frozen definitions, contamination controls, validated scoring, and a release boundary that does not leak the hidden holdout.
## What may be published here
- evaluator methodology
- schemas and validators
- scoring code when safe
- contamination-check methodology
- public example items
- evaluator version history
- negative results and known confounds
## What remains private
- hidden V1.8 holdout items
- sequestered evaluation material
- private prompts/contracts
- any artifact whose release would invalidate future measurement
Project: https://devim.org
Source: https://devim.org
|