NebulaExp-8B: Transparent Post-Training Pipeline for LLM Alignment.

Post-training alignment is critical for equipping large language models (LLMs) with strong reasoning abilities and the capacity to follow human preferences. However, many existing works lack detailed information on data construction, filtering rules, and training recipes, which hinders reproducibility and efficient optimization for the broader community. NebulaExp-8B addresses this transparency gap by offering a comprehensive, ablation-driven post-training pipeline built on the Qwen3-8B-base model. It details the curation of a 3.84M multi-source SFT sample corpus and a 200K verifiable RL candidate pool, alongside an end-to-end data processing stack that includes response distillation, multi-dimensional cross-verification filtering, fine-grained difficulty grading, task classification, and diversity-aware sampling. The pipeline develops two orthogonal model branches: a general instruct model and a complex reasoning-specialized model. For the Instruct branch, a three-stage optimized supervised fine-tuning (SFT) approach, NebulaExp-Ins-SFT, improved benchmark scores from 55.01 to 60.99, further elevated to 61.85 with GRPO reinforcement learning. For the Reasoning branch, medium-difficulty GRPO RL boosted scores from 73.88 to 75.17. The research also systematically investigates single-teacher and multi-teacher OPD (MOPD), showing that MOPD, with only 10K samples, can outperform RL baselines by fusing domain-specialist teachers. This work provides a fully reproducible recipe and dissects capability trade-offs across instruction adherence, mathematical reasoning, code generation, and general knowledge.