Hugging Face's logo

RedHatAI
/
GLM-5.3-speculator.dspark

Safetensors
speculators
speculative-decoding
dspark
long-context
custom_code
Model card Files
xet
 Community
2

Will the GLM-5.3 on-policy training dataset (~1.81M samples) be released?

#1
by Frozen7771 - opened
Discussion
Frozen7771

Hi @mgoin and the Red Hat AI team, thanks for releasing this speculator.

The model card and training_state.json indicate the draft was warm-started and trained for 1 epoch (88,226 steps) on ./data.jsonl, roughly 1.81M sequences at 8192 seq-len. For GLM-5.2 you published the regenerated corpus as mgoin/open-perfectblend-glm5.2-regen, which has been extremely useful.

Is there a plan to publish the GLM-5.3 on-policy regeneration dataset in the same way? If a public release isn't possible, would you be able to share the recipe: prompt sources / mix ratios, the regeneration setup (per-turn on-policy like the 5.2 set? reasoning_effort? max tokens per turn?), and how the 1.81M samples were filtered?

Thanks!
shanjiaz
 Red Hat AI org

Hey @Frozen7771 , happy to share the regen recipe!

python3 scripts/response_regeneration/script.py \
  --endpoint http://127.0.0.1:8000/v1/chat/completions \
  --dataset open-perfectblend \
  --concurrency 128 \
  --max-tokens 14000 \
  --reasoning-effort-cycle low,high,max \
  --outfile ./production/open-perfectblend_GLM-5.3.jsonl \
  --resume
Frozen7771

Thanks @shanjiaz , this is exactly what we needed — same open-perfectblend prompts as the 5.2 run makes it straightforward to reproduce against our own GLM-5.3 deployment.

Three quick follow-ups, if you have a moment:

  1. Is there any plan to publish the resulting ~1.81M-sample dataset itself (as you did for GLM-5.2 with mgoin/open-perfectblend-glm5.2-regen)? Even a "no, licensing" is useful to know, so we don't wait on it.
  2. Is the 1,811,739 figure the count of packed 8192-token sequences after prepare_data.py, rather than the number of regenerated conversations? Just so we size our own run correctly.
  3. Does --reasoning-effort-cycle rotate per request (roughly 1/3 each of low/high/max), or per pass over the dataset?

Thanks again.
shanjiaz
 Red Hat AI org

@Frozen7771 No worries!

  1. We might not publish the regen dataset any time soon. I'm sure if you have a production need RedHatAI would have something for you tho!
  2. That is the number of regenerated conversations and also the number of prepared Arrow examples after prepare_data.py.
  3. This command is compatible with Speculators latest release, and yes meaning roughly 1/3 of each effort. Recently we landed a feature on main to allow passing in a dict of reasoning effort distribution {low: 0.2, high:0.2, max:0.6}.

Happy to answer more questions on vllm slack https://join.slack.com/t/vllm-dev/shared_invite/zt-4a7elr2d3-b8uOC~GfdSZpooOsBKF9Rw My name is Helen Zhao. 🫡

Frozen7771

Thanks @Helen , that settles it — we'll reproduce with the open-perfectblend prompts and the effort distribution feature rather than wait on the dataset. Appreciate the Slack invite, will follow up there with anything further. 🫡

Sign up or log in to comment