Open Model Lab

Datasets

The frozen 25-task July suite, the planned eval suite v2, and the prepared SFT pilot support the next controlled training experiment.

Purpose

The project needs dataset cards because behavior changes cannot be interpreted without knowing which data was used for evals, SFT, DPO, agent tasks, or safety checks.

Eval leakage control is mandatory: training data must not contain eval tasks.

Published evaluation dataset

Name
july_eval_v1
Version
v1
Status
Published diagnostic baseline
Tasks
25
Categories
5, balanced at 5 tasks each
Quality boundary
All 25 task records are marked draft

The suite is a small diagnostic baseline, not a statistically representative benchmark. Its content hash is 38c646401b74f92deac04c5e225dfeabd822f72c3e30ce52dfaa24d817375760.

Prepared SFT pilot

Dataset
august_sft_v1
Prepared
2026-09-24
Training / validation
200 / 40
Smoke subset
24 training examples

A small procedural dataset for instruction following, structured output, short context-grounded answers, and arithmetic presentation. Production training follows exact-base eval suite v2 runs and a validated local smoke update/save/reload; no target month is committed.

The dataset keeps its original version identifier. Its content hash is bdc8d479735f7ba0eb6c195105127b3ad523c79d92ff92b74ee007f53a1a9de4.

Planned later datasets

  • eval_v2
  • instruction_v2_filtered
  • preference_v1
  • agent_tasks_v1
  • safety_eval_v1
  • ui_eval_v1
  • pretraining_ladder_v1

Dataset card fields

  • name
  • purpose
  • source
  • license/usage constraints
  • size
  • categories
  • filtering steps
  • deduplication
  • contamination/leakage checks
  • risks
  • version