project string | artifact string | registry_works int64 | sequestered_works int64 | derivation_works int64 | training_side_passages int64 | source_tokens int64 | book_grounded_support_review dict | abstention_review dict | release_boundary string |
|---|---|---|---|---|---|---|---|---|---|
DEVIM | Grounded Turkish public research frame | 593 | 119 | 474 | 147,307 | 50,863,441 | {
"supported": 109,
"reviewed": 110,
"unsupported": 0
} | {
"valid": 30,
"reviewed": 30
} | methodology_and_public_aggregates_only |
DEVİM Grounded Turkish
This is a methodology and public-evidence repository, not a release of the underlying rights-restricted Turkish corpus.
DEVİM's grounded-data program was created to reduce shortcut learning and weak transfer by linking supervision to source evidence, preserving provenance, separating training material from sequestered evaluation material, and explicitly testing abstention when an answer is not supported.
Verified source frame
- Authorized registry works: 593
- Permanently sequestered works: 119
- Derivation-pool works: 474
- Training-side passages: 147,307
- Source tokens in this frame: 50,863,441
The 119 sequestered works are not part of the derivation pool.
Public validation signals
Safe aggregate review results include:
- Book-grounded support review: 109 / 110 supported
- Unsupported in that review: 0
- Abstention review: 30 / 30 valid
These figures describe bounded validation samples; they are not a claim that every generated item is correct.
What is intentionally not released
- rights-restricted source text
- full derived training corpus
- sequestered works or hidden holdouts
- private generation prompts/contracts
- exact strategic generation recipes
- unpublished training checkpoints
The goal is selective reproducibility: publish enough structure and evidence to inspect the research method without destroying evaluation integrity or redistributing restricted material.
Project: https://devim.org
Source: https://devim.org
- Downloads last month
- 40