🧪 Model Behavior Incident Log

Collapse Symptoms After Annealing + SFT (Gemma-3-270M-TW)

• Gemma-3-270M (base) → CPT → Dialogue-Annealing → SFT
• Training performed with LLaMA-Factory on multiple Traditional Chinese datasets.

• what happened
• why the model collapsed
• representative failure outputs
• suspected root causes
• mitigation steps and future plans

📉 Summary of the Incident

• repetitive loops
• meaningless vocalization (e.g., hihihihi...)
• empty or nearly empty responses
• echoing the user query
• hallucinated list expansions
• loss of chat-template consistency
• collapse-like distribution drift

🔍 Representative Failure Outputs

1. Character-level repetition

2. Garbled symbols / noise

3. Semantic drift + self-looping

是什麼？
有，他們都是那個。
你們都 是那個。
我 還是在那個裡頭，
所以，沒了，沒有，沒了。。
我也不知道。
我知道。
知道。
知道了。

4. Query echoing + hallucinated QA scaffolding

請用一句話解釋地球為什麼會自轉。
- 根據物理學，什麼是「氣體」？
- 什麼是「磁場」？

5. Baseline sanity check (pre-annealing) behaves normally

🧭 Suspected Root Causes

1. Annealing Dataset Format Drift

user: ...
assistant: ...

2. Overwriting Instruction Alignment

3. SFT Unable to Recover

🛠 Actions Taken

• Rebuilt CPT dataset (2.4M samples, 70/30 mix)
• Adopted stable LR tail (cosine_with_min_lr)
• Will rebuild annealing data with strict chat format
• Full SFT pipeline will be repeated after clean CPT

🔄 Next Steps

1. Complete clean CPT
2. Rebuild dialogue annealing
3. Annealing from min_lr → 0
4. Re-run SFT
5. Compare ablation paths
6. Add regression tests

🧭 Closing Note