aabbdev commited on
Commit
6b20acf
·
verified ·
1 Parent(s): 45b1e2b

Document dense-100m configuration

Browse files
Files changed (1) hide show
  1. README.md +14 -0
README.md CHANGED
@@ -19,6 +19,12 @@ configs:
19
  path: "data/train/*"
20
  - split: validation
21
  path: "data/validation/*"
 
 
 
 
 
 
22
  ---
23
 
24
  <div align="center">
@@ -35,6 +41,10 @@ One deterministic corpus. Three training views: **pretraining · SFT · on-polic
35
  |:--|:--|:--|:--|
36
  | **10,000,000 rows** | **36 families · M1–M11** | **7.80 GiB · 41 Parquet shards** | **9,949,600 train · 50,400 validation** |
37
 
 
 
 
 
38
  StateBench teaches models to execute finite state programs rather than recall facts. A typed program is executed by a deterministic reference evaluator; only then is it rendered into varied textual surfaces.
39
 
40
  ## Load it
@@ -42,11 +52,15 @@ StateBench teaches models to execute finite state programs rather than recall fa
42
  from datasets import load_dataset
43
 
44
  ds = load_dataset("aabbdev/StateBench-v1", split="train", streaming=True)
 
45
 
46
  pretrain = ds.select_columns(["text"])
47
  sft = ds.select_columns(["prompt", "completion"])
48
  sdft = ds.select_columns(["prompt", "privileged_context"])
49
 
 
 
 
50
  ## One corpus, three views
51
 
52
  | Regime | Student input | Training signal |
 
19
  path: "data/train/*"
20
  - split: validation
21
  path: "data/validation/*"
22
+ - config_name: dense-100m
23
+ data_files:
24
+ - split: train
25
+ path: "data/dense-100m/train/*"
26
+ - split: validation
27
+ path: "data/dense-100m/validation/*"
28
  ---
29
 
30
  <div align="center">
 
41
  |:--|:--|:--|:--|
42
  | **10,000,000 rows** | **36 families · M1–M11** | **7.80 GiB · 41 Parquet shards** | **9,949,600 train · 50,400 validation** |
43
 
44
+ | Dense extension | Physical rows | Density | Storage |
45
+ |:--|:--|:--|:--|
46
+ | **100,000,000 episodes** | **25,000,000 packed rows** | **162.01 units/1k chars** | **94.78 GiB · 251 shards** |
47
+
48
  StateBench teaches models to execute finite state programs rather than recall facts. A typed program is executed by a deterministic reference evaluator; only then is it rendered into varied textual surfaces.
49
 
50
  ## Load it
 
52
  from datasets import load_dataset
53
 
54
  ds = load_dataset("aabbdev/StateBench-v1", split="train", streaming=True)
55
+ dense = load_dataset("aabbdev/StateBench-v1", "dense-100m", split="train", streaming=True)
56
 
57
  pretrain = ds.select_columns(["text"])
58
  sft = ds.select_columns(["prompt", "completion"])
59
  sdft = ds.select_columns(["prompt", "privileged_context"])
60
 
61
+ The default configuration stores text directly. The dense-100m configuration omits that
62
+ redundant physical column; derive it from prompt and completion when using causal pretraining.
63
+
64
  ## One corpus, three views
65
 
66
  | Regime | Student input | Training signal |