mgoin/GLM-5.2-FP8-magpie-ultrachat

GLM-5.2-FP8 Regenerated Responses (Magpie + UltraChat mix)

A combined instruction-response dataset of 507,864 single-turn conversations. The prompts are drawn from two public instruction datasets; the responses were freshly regenerated with zai-org/GLM-5.2-FP8.

It was built as on-policy distillation data for training speculative-decoding drafts (DFlash / DSpark) for GLM-5.2 — i.e. so the draft learns from GLM-5.2's own output distribution — but it is a general-purpose GLM-5.2 response set.

Composition

source	prompts from	rows
ultrachat	HuggingFaceH4/ultrachat_200k (train_sft, first-turn prompt)	207,864
magpie	Magpie-Align/Magpie-Llama-3.1-Pro-300K-Filtered (instruction)	300,000
total		507,864

How it was made

Each source prompt was sent to a vLLM server hosting zai-org/GLM-5.2-FP8 and the model's completion was recorded. No responses were filtered out; every prompt has exactly one response.

• Model: zai-org/GLM-5.2-FP8 (served with vLLM, tensor-parallel across 8 GPUs)
• API: OpenAI-compatible /v1/chat/completions, one user turn per prompt
• Sampling: vLLM server defaults, max_tokens=1024
• Deduplicated by prompt id; a resume/retry loop guaranteed one successful response per prompt.

Important characteristics (read before using)

• Responses contain GLM-5.2's inline reasoning. GLM-5.2 emits a <think>…</think> reasoning trace before its answer, and the server ran without a reasoning parser, so the reasoning is kept inline in the response text. A complete </think> block is present in ~57% of ultrachat and ~34% of magpie responses.
• Many responses are truncated at 1024 tokens. Because GLM-5.2's reasoning traces are long and generation was capped at max_tokens=1024, most responses hit the length limit (finish_reason=length: ~74% of ultrachat, ~90% of magpie) and are cut off — sometimes mid-reasoning. This is acceptable for speculative-decoding distillation (the goal is the target model's token distribution over realistic prefixes) but makes the data unsuitable as-is for supervised fine-tuning on complete answers. Filter on finish_reason == "stop" if you need complete responses.

Format

ShareGPT-style JSONL. One example:

{
  "id": "sample_1234",
  "source": "ultrachat",
  "conversations": [
    {"from": "human", "value": "<prompt>"},
    {"from": "gpt",   "value": "<GLM-5.2-FP8 response, may include <think>…</think>>"}
  ],
  "finish_reason": "length"
}

field	description
id	stable id (source uuid, or sample_<index> for UltraChat)
source	ultrachat or magpie
conversations	[{from: human, value}, {from: gpt, value}]
finish_reason	stop (natural end) or length (hit max_tokens=1024)

Licensing / provenance

This is a derivative work. Downstream use should respect all upstream terms:

• Prompts — ultrachat_200k (MIT) and Magpie-Llama-3.1-Pro-300K-Filtered (Llama 3.1 Community License).
• Responses — generated by zai-org/GLM-5.2-FP8 (see the model's license).

Tagged license: other because it combines sources under different terms.