cbarkinozer/Occam-Turkish-Response-SFT

Occam — Turkish Response SFT

A small, human-reviewed Turkish response-style dataset curated by the author, cbarkinozer, for the Occam adapter.

The repository owner selected Apache-2.0; see LICENSE. This private upload does not attest completion of historical provenance/provider-terms review.

Purpose and composition

Human-reviewed, synthetic/AI-assisted Turkish conversations used for Occam LoRA response-style SFT on Qwen3.5-4B. 104 training rows and 20 development rows. The initial80/20 selection was followed by24 additional training scenarios; the released 104/20 split is not an 80/20 percentage split. Twelve independent project control items and all external CETVEL inputs/references are excluded. These are training targets, not labels certifying factual correctness in every case.

Desired behaviors: immediate answers and requested formatting; accurate extraction; faithful shortening preserving uncertainty/negation/conditions; useful explanations; clarification only when needed; no unsupported causes, motives or guarantees; Turkish character/word/suffix/tense precision; no fabricated model feelings, preferences or memories; neutral acknowledgment without reflexive advice/disclaimers. Appropriate longer answers are allowed: brevity is not the sole objective.

Creation and provenance

Questions were selected/curated from historical project inventories and new project-agent-authored scenarios. Original Qwen3.5-4B generations informed minimal edits, then a human reviewed/rewrote targets, sometimes assisted by a personal AI assistant. Later core target repairs and24 reviewed independent additions formed the released dataset. This is not naturally occurring human conversation and not wholly human-authored. The exact personal-assistant provider/model and terms must be supplied by the owner; some historical synthetic provenance is incomplete. Do not substitute a guessed teacher model or claim every sample derives from the same generator. source-origins.json preserves available origin/category counts; immutable original experiment files carry fuller internal provenance and review receipts.

Schema and supervision

JSONL with id, category, subtype, scenario_group_id, origin, source_id, evaluation_only, and conversational messages (role/content). A final assistant message contains the exact reviewed target. Earlier assistant turns are context, NOT training targets under the recorded final-assistant-only loss policy. Review criteria, original generations, user/review identifiers, annotations and internal file paths are not included in the public training messages.

from datasets import load_dataset
dataset = load_dataset("cbarkinozer/Occam-Turkish-Response-SFT")
print(dataset["train"][0]["messages"])

Apply the pinned model's native non-thinking chat template. Do not train raw user: strings or include the annotation rubric as instructions. Hashes of the source and clean release files are in release-manifest.json; serialization differs from internal provenance-rich rows, but conversation/target content is preserved.

Split and leakage limitations

Train/validation exact/near/scenario checks found no identified item-level leak; retained task-family similarities and generic short answers are documented in SEMANTIC-LEAKAGE-REVIEW.md in the code repository. Audits do not prove zero semantic overlap. The20 validation prompts were repeatedly used during model/data selection: they are a development set, not a clean final evaluation. One reference, me-089, was disputed and preserved; use recorded sensitivity analysis rather than silently treating all references as authoritative. Do not add CETVEL questions, references or paraphrases to these training rows. External evaluation-source licensing is separate from this dataset's license.

Limitations, privacy and rights

Small, curated, uneven task distribution; many everyday synthetic situations. No evidence that similar token frequencies prevent catastrophic forgetting. Human review can miss unnatural Turkish, incorrect reasoning and unsupported claims. Historical generation cannot be reproduced exactly from all surviving metadata. Review every released row for privacy/third-party content before public upload; this draft does not assert a completed legal or exhaustive privacy audit. Dataset license: Apache-2.0, selected by the repository owner. Historical provenance and provider-terms review remain incomplete; a license selection alone does not establish all underlying publication rights.