qianlihuang commited on
Commit
ed368e1
·
verified ·
1 Parent(s): fd11787

Upload folder using huggingface_hub

Browse files
Files changed (3) hide show
  1. .gitattributes +1 -0
  2. README.md +53 -0
  3. muse_opb_onpolicy_100k.jsonl +3 -0
.gitattributes CHANGED
@@ -58,3 +58,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
58
  # Video files - compressed
59
  *.mp4 filter=lfs diff=lfs merge=lfs -text
60
  *.webm filter=lfs diff=lfs merge=lfs -text
 
 
58
  # Video files - compressed
59
  *.mp4 filter=lfs diff=lfs merge=lfs -text
60
  *.webm filter=lfs diff=lfs merge=lfs -text
61
+ muse_opb_onpolicy_100k.jsonl filter=lfs diff=lfs merge=lfs -text
README.md ADDED
@@ -0,0 +1,53 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ task_categories:
4
+ - text-generation
5
+ language:
6
+ - en
7
+ tags:
8
+ - speculative-decoding
9
+ - on-policy
10
+ - open-perfect-blend
11
+ - muse-glimmer
12
+ pretty_name: Muse Glimmer OPB 100K
13
+ size_categories:
14
+ - 100K<n<1M
15
+ ---
16
+
17
+ # Muse Glimmer OPB 100K
18
+
19
+ On-policy OpenPerfectBlend training data used for [DaoCloud/Muse-Glimmer-30B-DSpark](https://huggingface.co/DaoCloud/Muse-Glimmer-30B-DSpark).
20
+
21
+ Prompts are sampled from [mlabonne/open-perfectblend](https://huggingface.co/datasets/mlabonne/open-perfectblend), and assistant turns are regenerated on-policy with Muse Glimmer 30B.
22
+
23
+ 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.
24
+
25
+ | Reasoning strength | Conversations | Train-turn rows |
26
+ |---|---:|---:|
27
+ | low | 64,997 | 96,765 |
28
+ | medium | 10,000 | 14,829 |
29
+ | high | 19,991 | 29,957 |
30
+ | xhigh | 4,996 | 7,349 |
31
+ | **Total** | **99,984** | **148,900** |
32
+
33
+ ## Schema
34
+
35
+ | Field | Description |
36
+ |---|---|
37
+ | `id` | Unique train-turn ID |
38
+ | `primary_id` | Source conversation ID |
39
+ | `input_ids` | Tokenized on-policy conversation through the train turn |
40
+ | `loss_mask` | Token-level training mask |
41
+ | `reasoning_strength` | `low`, `medium`, `high`, or `xhigh` |
42
+
43
+ Every row has trainable tokens. Row IDs are unique, source conversation sets are disjoint, and generation ordinals are contiguous.
44
+
45
+ ## File
46
+
47
+ ```text
48
+ muse_opb_onpolicy_100k.jsonl
49
+ 1,646,393,887 bytes
50
+ SHA-256 70d68dff971d1fe00d90a356b1a1eaeac845f9a3cc2069ce525043312444439a
51
+ ```
52
+
53
+ This release contains tokenized training examples. Use the Muse Glimmer 30B tokenizer when inspecting or converting `input_ids` back to text.
muse_opb_onpolicy_100k.jsonl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:70d68dff971d1fe00d90a356b1a1eaeac845f9a3cc2069ce525043312444439a
3
+ size 1646393887