metadata
license: apache-2.0
task_categories:
- text-generation
language:
- en
tags:
- speculative-decoding
- on-policy
- open-perfect-blend
- muse-glimmer
pretty_name: Muse Glimmer OPB 100K
size_categories:
- 100K<n<1M
Muse Glimmer OPB 100K
On-policy OpenPerfectBlend training data used for DaoCloud/Muse-Glimmer-30B-DSpark.
Prompts are sampled from mlabonne/open-perfectblend, and assistant turns are regenerated on-policy with Muse Glimmer 30B.
The dataset contains 99,984 successfully generated conversations and 148,900 train-turn rows. Responses were regenerated with Muse Glimmer 30B at four reasoning strengths.
| Reasoning strength | Conversations | Train-turn rows |
|---|---|---|
| low | 64,997 | 96,765 |
| medium | 10,000 | 14,829 |
| high | 19,991 | 29,957 |
| xhigh | 4,996 | 7,349 |
| Total | 99,984 | 148,900 |
Schema
| Field | Description |
|---|---|
id |
Unique train-turn ID |
primary_id |
Source conversation ID |
input_ids |
Tokenized on-policy conversation through the train turn |
loss_mask |
Token-level training mask |
reasoning_strength |
low, medium, high, or xhigh |
Every row has trainable tokens. Row IDs are unique, source conversation sets are disjoint, and generation ordinals are contiguous.
File
muse_opb_onpolicy_100k.jsonl
1,646,393,887 bytes
SHA-256 70d68dff971d1fe00d90a356b1a1eaeac845f9a3cc2069ce525043312444439a
This release contains tokenized training examples. Use the Muse Glimmer 30B tokenizer when inspecting or converting input_ids back to text.