sidbaines commited on
Commit
d70e5ff
·
verified ·
1 Parent(s): 96df04f

rollouts/gemma4_26b_a4b_190m: provenance sidecars

Browse files

The dose study's two rollout archives were copied verbatim (already gzipped at source, nothing dropped) and so arrived without the .meta.json every other archive here carries. These record the source, the row count, the fields present, and that the stored bytes reproduce the source's own content key -- and state explicitly that, unlike the rlvr-gemma4-26b archives, no fields were removed.

rollouts/gemma4_26b_a4b_190m/charter-direct/raw_rollouts.rank-0.jsonl.meta.json ADDED
@@ -0,0 +1,44 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "source_repo": "sidbaines/scimt-dispatch-gemma4-26b-charter-190m-graft-v1",
3
+ "source_path": "rollouts/charter-direct/raw_rollouts.rank-0.jsonl.gz",
4
+ "verbatim_copy": true,
5
+ "dropped_fields": [],
6
+ "declared_bytes": 8306838,
7
+ "stored_bytes": 8306838,
8
+ "sha256": "5cb00e9833555596f782045b826e6d434f8f2b5b72a62834fd67d30a201c007d",
9
+ "source_key_expected": "5cb00e9833555596f782045b826e6d434f8f2b5b72a62834fd67d30a201c007d",
10
+ "source_key_form": "lfs sha256",
11
+ "sha256_match": true,
12
+ "rows": 49152,
13
+ "fields": [
14
+ "channel_close_count",
15
+ "channel_open_count",
16
+ "completion",
17
+ "completion_ids",
18
+ "completion_length",
19
+ "completion_raw_text",
20
+ "completion_truncated",
21
+ "episode",
22
+ "episode_id",
23
+ "format_valid",
24
+ "log_extra",
25
+ "log_metric",
26
+ "native_boundary_valid",
27
+ "parser_json",
28
+ "parser_labelled_records",
29
+ "parser_natural",
30
+ "parser_unsafe",
31
+ "parser_valid",
32
+ "prompt",
33
+ "prompt_template_id",
34
+ "reward",
35
+ "reward_call",
36
+ "runs_correct",
37
+ "runs_total",
38
+ "selection_key",
39
+ "semantic_correct",
40
+ "truncated"
41
+ ],
42
+ "note": "Copied byte-for-byte from the source, which was already gzipped. UNLIKE the rlvr-gemma4-26b archives beside it, nothing was dropped -- those had trainer_state and completion_ids removed; this one is complete as sampled.",
43
+ "recorded": "2026-09-11"
44
+ }
rollouts/gemma4_26b_a4b_190m/control-direct/raw_rollouts.rank-0.jsonl.meta.json ADDED
@@ -0,0 +1,44 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "source_repo": "sidbaines/scimt-dispatch-gemma4-26b-charter-190m-graft-v1",
3
+ "source_path": "rollouts/control-direct/raw_rollouts.rank-0.jsonl.gz",
4
+ "verbatim_copy": true,
5
+ "dropped_fields": [],
6
+ "declared_bytes": 8569182,
7
+ "stored_bytes": 8569182,
8
+ "sha256": "bb71421de65a4e5059c954fa6cd4be3c2fa548acc8589052eda9d20eae31d04b",
9
+ "source_key_expected": "bb71421de65a4e5059c954fa6cd4be3c2fa548acc8589052eda9d20eae31d04b",
10
+ "source_key_form": "lfs sha256",
11
+ "sha256_match": true,
12
+ "rows": 49152,
13
+ "fields": [
14
+ "channel_close_count",
15
+ "channel_open_count",
16
+ "completion",
17
+ "completion_ids",
18
+ "completion_length",
19
+ "completion_raw_text",
20
+ "completion_truncated",
21
+ "episode",
22
+ "episode_id",
23
+ "format_valid",
24
+ "log_extra",
25
+ "log_metric",
26
+ "native_boundary_valid",
27
+ "parser_json",
28
+ "parser_labelled_records",
29
+ "parser_natural",
30
+ "parser_unsafe",
31
+ "parser_valid",
32
+ "prompt",
33
+ "prompt_template_id",
34
+ "reward",
35
+ "reward_call",
36
+ "runs_correct",
37
+ "runs_total",
38
+ "selection_key",
39
+ "semantic_correct",
40
+ "truncated"
41
+ ],
42
+ "note": "Copied byte-for-byte from the source, which was already gzipped. UNLIKE the rlvr-gemma4-26b archives beside it, nothing was dropped -- those had trainer_state and completion_ids removed; this one is complete as sampled.",
43
+ "recorded": "2026-09-11"
44
+ }