| --- |
| pretty_name: Stage 3 Final Mixture — 50% CoT Compression |
| license: apache-2.0 |
| task_categories: |
| - text-generation |
| --- |
| |
| # Stage 3 Final Mixture — 50% CoT Compression |
|
|
| This is a deterministic capability-preserving rewrite of |
| `leonli66/stage3-final-mixture` for LCLM Stage-3 post-training. |
|
|
| Only the `reasoning_data` and `dolci_think` subsets change. Their |
| `compression_prompt` is the ordinary `prompt`. A deterministic 50% arm keeps |
| the complete assistant target as ordinary SFT; the other arm wraps the inferred |
| reasoning prefix in `<|memory_start|>...<|memory_end|>` while keeping the final |
| answer trainable. All non-reasoning rows are unchanged. |
|
|
| ## Counts |
|
|
| - Total rows: 20,326,114 |
| - Reasoning rows: 5,407,421 |
| - Assigned CoT compression: 2,702,448 (49.98%) |
| - Assigned ordinary uncompressed SFT: 2,704,973 (50.02%) |
| - Effective CoT-compressed rows: 2,698,545 |
| - Compression-assigned rows too short to split: 3,903 |
|
|
| ## Boundary rules |
|
|
| Priority is: explicit analysis tags; final-answer markers; a final boxed answer; |
| an answer line; the final balanced fenced block; a GSM-style final `####` line; |
| then a fallback that keeps the final 128 |
| `Qwen/Qwen3-4B-Instruct-2507` tokens uncompressed. |
|
|
| The transformation is text-preserving: deleting the two LCLM memory tags from |
| every rewritten target reconstructs the original target exactly. All 2,033 |
| Parquet shards were validated against the source. |
|
|
