GLM-5.2-FP8-nemotron-codealpaca

# GLM-5.2-FP8-nemotron-codealpaca Training data for [UCloud-AILab/GLM-5.2-FP8-DFlash](https://modelscope.cn/models/UCloud-AILab/GLM-5.2-FP8-DFlash), a DFlash speculative-decoding drafter for [GLM-5.2-FP8](https://modelscope.cn/models/ZhipuAI/GLM-5.2-FP8). A mix of code / math / chat prompts from two public instruction datasets (see Composition); **all assistant responses are regenerated by GLM-5.2-FP8** so the targets match the verifier's own output distribution — the data recipe specified in the [DFlash paper](https://arxiv.org/abs/2602.06036) (Appendix A.1). - **800,022 single-turn conversations**, English-dominant - Generation: `temperature=1.0`, `max_tokens=4096`, model default chat template - The published files match the drafter training input after credential-like strings in selected user and assistant turns were redacted with `*`; see Corrections. Of the 800,022 rows, 788,966 (~98.6%) passed length/validity preprocessing (responses truncated to 3,072 tokens) and entered the final 6-epoch training run ## Composition | Source | Prompt origin | Rows | |--------|---------------|------| | nemotron-pt-v2 | [nvidia/Nemotron-Post-Training-Dataset-v2](https://huggingface.co/datasets/nvidia/Nemotron-Post-Training-Dataset-v2) (chat 365,533 / math 239,467 / code 175,000) | 780,000 | | codealpaca | [CodeAlpaca-20k](https://modelscope.cn/datasets/AI-ModelScope/CodeAlpaca-20k) | 20,022 | | **Total** | | **800,022** | ## Characteristics - Responses contain **no `<think>` reasoning blocks** (0% in a 20K sample of nemotron rows and in all codealpaca rows) — targets are plain final answers - Small long-response tail: ~7.5% of nemotron and ~2.5% of codealpaca responses exceed 12,000 characters and may be truncated at `max_tokens=4096` - `status == "success"` on every row; no failed or empty generations ## Corrections ### 1. CodeAlpaca prompts are incomplete for 10,258 of the 20,022 CodeAlpaca rows The conversion used `instruction` but omitted CodeAlpaca's optional `input` field. Consequently, 10,258 rows with non-empty `input` contain shortened prompts and responses generated from those prompts. The remaining 9,764 CodeAlpaca rows had empty `input` fields and are unaffected. ### 2. The generation temperature was 1.0 An earlier version of this card incorrectly reported `temperature=0.8`. The actual generation setting was `temperature=1.0`; `max_tokens=4096` was reported correctly. ### 3. Credential-like strings were redacted in both roles Redaction masks occur in user turns of 165 rows and assistant turns of 122 rows. These counts are role-specific and the two sets overlap; no distinct-row total is claimed. The basis of the earlier figure of 28 cannot be reconstructed reliably, so it is withdrawn. **Status:** The published data files are unchanged and will not be silently replaced. A separately named corrected release is in preparation. ## Format One JSON object per line: ```json { "id": "…", "conversations": [ {"role": "user", "content": "…"}, {"role": "assistant", "content": "…"} ], "status": "success", "domain": "chat | code | math (nemotron rows) or codealpaca", "source": "nemotron-pt-v2 | codealpaca", "char_len": 5796 } ``` `char_len` is the total character count across all turns of the conversation (sampled median ≈ 5,800). Files: `data/train-00000-of-00012.jsonl` … `data/train-00011-of-00012.jsonl` — a single `train` split. Rows are stored grouped by source/domain (not shuffled); shuffle at load time for training use. The per-row `source` field distinguishes the two prompt origins. ## Load ```python from modelscope.msdatasets import MsDataset ds = MsDataset.load("Jessiewei7/GLM-5.2-FP8-nemotron-codealpaca") ``` Also mirrored on Hugging Face: [JessieWei/GLM-5.2-FP8-nemotron-codealpaca](https://huggingface.co/datasets/JessieWei/GLM-5.2-FP8-nemotron-codealpaca). ## Licensing Prompts inherit their source licenses (Nemotron-Post-Training-Dataset-v2: CC-BY-4.0; CodeAlpaca-20k: CC-BY-4.0). Assistant responses are machine-generated by GLM-5.2-FP8 (MIT). The compilation is released under CC-BY-4.0. ## Citation If you use this dataset, please cite: ```bibtex @misc{wei2026glm52regen, title = {GLM-5.2-FP8 Regenerated Nemotron-CodeAlpaca Dataset}, author = {Wei, Xiaojun and {UCloud AILab}}, year = {2026}, howpublished = {\url{https://huggingface.co/datasets/JessieWei/GLM-5.2-FP8-nemotron-codealpaca}} } ``` Prompts derive from Nemotron-Post-Training-Dataset-v2 and CodeAlpaca-20k — please credit them per the CC-BY terms above. For the drafter trained on this data (and the DFlash method citation), see [UCloud-AILab/GLM-5.2-FP8-DFlash](https://modelscope.cn/models/UCloud-AILab/GLM-5.2-FP8-DFlash).