Clarification on whether Qwen3.5-9B-Base received any post-training or assistant-aligned mid-training
Hello Qwen team,
Thank you for releasing Qwen3.5-9B-Base. I'm evaluating it for foundation-model R&D where the starting checkpoint needs to be as close as possible to a clean base model. I hope to implement my own continued pretraining and full post-training stack, so I want to correctly understand which training stages were applied before this checkpoint was released. The model card describes this repository as the “pre-trained only model,” but the overview also lists “Training Stage: Pre-training & Post-training.” Could you clarify what this means for Qwen3.5-9B-Base specifically?
In particular, did the released 'Base' checkpoint receive any assistant-aligned or post-training stage before release, such as:
- supervised fine-tuning / instruction tuning
- assistant-response or chat tuning
- RLHF, RLAIF, DPO, GRPO, PPO, or other preference/RL optimization
- safety alignment
- reasoning or chain-of-thought tuning
- tool-use or agentic-behavior training
- chat-template behavioral tuning beyond tokenizer/control-token support
- continued pretraining or midtraining on instruction, chat, assistant-response, tool-use, or reasoning-solution data
I understand that continued pretraining or midtraining on raw/non-interactive corpora such as code, papers, math, documents, web text, or multilingual text may still be considered part of pretraining. My concern is specifically whether any deliberate assistant-aligned, instruction-style, preference-based, tool-use, or reasoning-solution training was applied to the 'Base' checkpoint, since that would confound independent post-training stack R&D.
I do not need proprietary details about the dataset or training recipe. A high-level clarification would be enough: Is Qwen3.5-9B-Base suitable as a starting checkpoint for independent continued pretraining and a fully custom post-training pipeline, assuming I need no prior deliberate assistant alignment, preference optimization, or instruction tuning to have been applied?
Thank you very much for clarifying!