Siyuc commited on
Commit
4ec391c
·
verified ·
1 Parent(s): 3acbd0a

Add original and prefix-query benchmark result snapshots

Browse files
benchmarks/20260920-151814_paper2604/README.md ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Paper2604 benchmark-result snapshots
2
+
3
+ Snapshot published 2026-09-20T15:18:17.531373+08:00 (Asia/Shanghai).
4
+
5
+ | File | Scope | Checkpoints | Status at snapshot |
6
+ |---|---|---:|---|
7
+ | [original/ALL_RESULTS.json](original/ALL_RESULTS.json) | Original20-task suite;68754 examples per full evaluation |22|{"succeeded": 21, "running": 1}|
8
+ | [prefix_query/ALL_RESULTS.json](prefix_query/ALL_RESULTS.json) | Four Variable Assignment prefix-query benchmarks;4000 examples per checkpoint |22|{"succeeded": 22}|
9
+
10
+ Original report updated: 2026-09-20T15:17:49.329692+08:00.
11
+ Prefix-query report updated: 2026-09-20T14:23:42.867303+08:00.
12
+
13
+ These are the two requested JSON files, copied byte for byte, not a merged24-task aggregate. Full checkpoint names, steps, per-task metrics, status and provenance are retained. The prefix-query JSON also includes original scores for the four corresponding tasks for paired comparisons.
14
+
15
+ The original and prefix-query suites retain separate aggregates. A running row is incomplete; its partial task scores must not be interpreted as a full-suite result. This upload is a fixed snapshot, not a live report, and does not claim that an in-progress original-suite evaluation has finished.
16
+
17
+ Prefix-query examples preserve original assignments, gold answers and five-shot selections. Query replacement applies to demonstrations and the held-out question. Math uses `x=4`, Code `x = 4`; changed answer spacing can affect token-level loss. All evaluated checkpoints use the recorded native architecture-specific inference, FP32 and no truncation.
18
+
19
+ Evaluation implementation base: `be553afcfbbb9cb3425e24f788b16a5a0bca221c`. Prefix-query implementation and documentation published on GitHub in `414cc438e67f921df1fcd218dcd738b88c83c2e1`; live evaluations retained their pinned worktree. These artifacts contain results, not model weights or training-resume state.
20
+
21
+ `manifest.json` records SHA256 hashes and byte sizes for these JSON files and this README. It does not hash itself. Existing repository files and visibility are unchanged.
benchmarks/20260920-151814_paper2604/manifest.json ADDED
@@ -0,0 +1,42 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "created": "2026-09-20T15:18:17.538818+08:00",
3
+ "repository": "Siyuc/loopTF-workspace",
4
+ "destination": "benchmarks/20260920-151814_paper2604",
5
+ "source_reports": {
6
+ "original/ALL_RESULTS.json": {
7
+ "source": "/root/dl/runtime/script/loopTF-workspace/output/paper2604-all-20260920/ALL_RESULTS.json",
8
+ "updated": "2026-09-20T15:17:49.329692+08:00",
9
+ "checkpoints": 22,
10
+ "statuses": {
11
+ "succeeded": 21,
12
+ "running": 1
13
+ }
14
+ },
15
+ "prefix_query/ALL_RESULTS.json": {
16
+ "source": "/root/dl/runtime/script/loopTF-workspace/output/paper2604-prefix-query-20260920/ALL_RESULTS.json",
17
+ "updated": "2026-09-20T14:23:42.867303+08:00",
18
+ "checkpoints": 22,
19
+ "statuses": {
20
+ "succeeded": 22
21
+ }
22
+ }
23
+ },
24
+ "files": [
25
+ {
26
+ "path": "README.md",
27
+ "size": 2053,
28
+ "sha256": "6d2e5de6172cda97982cf5c8d69a3c4eb9c4ebf5cacde28323cf6d8f206c5dce"
29
+ },
30
+ {
31
+ "path": "original/ALL_RESULTS.json",
32
+ "size": 231195,
33
+ "sha256": "a42b1dca68eb1d0afee4a9418dc1239a13870eb7f24649b5a086bea65527d128"
34
+ },
35
+ {
36
+ "path": "prefix_query/ALL_RESULTS.json",
37
+ "size": 93586,
38
+ "sha256": "4c67176246293bd3461af9f54d4fdccc7c8fecb6f74bd679fad114cc2dd93b7b"
39
+ }
40
+ ],
41
+ "note": "Manifest does not hash itself."
42
+ }
benchmarks/20260920-151814_paper2604/original/ALL_RESULTS.json ADDED
The diff for this file is too large to render. See raw diff
 
benchmarks/20260920-151814_paper2604/prefix_query/ALL_RESULTS.json ADDED
@@ -0,0 +1,2707 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "updated": "2026-09-20T14:23:42.867303+08:00",
3
+ "suite": {
4
+ "schema_version": 1,
5
+ "protocol": "paper2604-variable-assignment-prefix-query-v1",
6
+ "protocol_status": "Custom paired query-format ablation; not part of the original paper2604 aggregate",
7
+ "upstreams": {
8
+ "paper2604": {
9
+ "repository": "kschwethelm/looped-lm-scaling",
10
+ "commit": "ad7d8f3055501e4c84ced7797f772dce2f8e2107",
11
+ "paper": "https://arxiv.org/pdf/2604.21106v3"
12
+ }
13
+ },
14
+ "tasks": [
15
+ {
16
+ "name": "var_assign_d0_math_prefix_query",
17
+ "display_name": "Variable Assignment D0 Math prefix query",
18
+ "source_task": "var_assign_d0_math",
19
+ "engine": "paper2604",
20
+ "task": "var_assign_d0_math_prefix_query",
21
+ "axis": "variable_assignment_prefix_query",
22
+ "bundle": "owned",
23
+ "fewshot": 5,
24
+ "task_type": "language_modeling",
25
+ "expected_examples": 1000,
26
+ "continuation_delimiter": "",
27
+ "metrics": [
28
+ "continuation_loss",
29
+ "ppl",
30
+ "accuracy"
31
+ ]
32
+ },
33
+ {
34
+ "name": "var_assign_d0_code_prefix_query",
35
+ "display_name": "Variable Assignment D0 Code prefix query",
36
+ "source_task": "var_assign_d0_code",
37
+ "engine": "paper2604",
38
+ "task": "var_assign_d0_code_prefix_query",
39
+ "axis": "variable_assignment_prefix_query",
40
+ "bundle": "owned",
41
+ "fewshot": 5,
42
+ "task_type": "language_modeling",
43
+ "expected_examples": 1000,
44
+ "continuation_delimiter": " ",
45
+ "metrics": [
46
+ "continuation_loss",
47
+ "ppl",
48
+ "accuracy"
49
+ ]
50
+ },
51
+ {
52
+ "name": "var_assign_d1_math_prefix_query",
53
+ "display_name": "Variable Assignment D1 Math prefix query",
54
+ "source_task": "var_assign_d1_math",
55
+ "engine": "paper2604",
56
+ "task": "var_assign_d1_math_prefix_query",
57
+ "axis": "variable_assignment_prefix_query",
58
+ "bundle": "owned",
59
+ "fewshot": 5,
60
+ "task_type": "language_modeling",
61
+ "expected_examples": 1000,
62
+ "continuation_delimiter": "",
63
+ "metrics": [
64
+ "continuation_loss",
65
+ "ppl",
66
+ "accuracy"
67
+ ]
68
+ },
69
+ {
70
+ "name": "var_assign_d1_code_prefix_query",
71
+ "display_name": "Variable Assignment D1 Code prefix query",
72
+ "source_task": "var_assign_d1_code",
73
+ "engine": "paper2604",
74
+ "task": "var_assign_d1_code_prefix_query",
75
+ "axis": "variable_assignment_prefix_query",
76
+ "bundle": "owned",
77
+ "fewshot": 5,
78
+ "task_type": "language_modeling",
79
+ "expected_examples": 1000,
80
+ "continuation_delimiter": " ",
81
+ "metrics": [
82
+ "continuation_loss",
83
+ "ppl",
84
+ "accuracy"
85
+ ]
86
+ }
87
+ ],
88
+ "defaults": {
89
+ "seed": 1337,
90
+ "fewshot_seed": 1234,
91
+ "max_per_task": 10000,
92
+ "dtype": "float32",
93
+ "context_policy": "error",
94
+ "model_max_length": 4096,
95
+ "batch_size": 1,
96
+ "wandb": false
97
+ }
98
+ },
99
+ "checkpoints": [
100
+ {
101
+ "name": "loop",
102
+ "checkpoint_name": "paper2604_loop2_476m_7b_online_FBW0_K3_adapter_norm_sharekv_20260918-043531/step_26703",
103
+ "checkpoint": "/taijifs_wzbd31/rentTest/hunyuan/kenanchen/loopTF-workspace/output/evaluation-readiness/checkpoints/loop/evaluation_native.pt",
104
+ "checkpoint_sha256": "570fcd66141a26ae73731d725daf907b2dcfa11b0e4f07f4cda1d3bdce71e372",
105
+ "step": 26703,
106
+ "architecture": "full_bandwidth_loop",
107
+ "parameter_count": 653383808,
108
+ "training_tokens": 7000031232,
109
+ "inference": {
110
+ "requested": "auto",
111
+ "resolved": "fused",
112
+ "full_bandwidth": true,
113
+ "prompt_traversals": 2,
114
+ "fused_prefill_passes": 1,
115
+ "continuation": "one feedback transition per real token",
116
+ "cache": "architecture-owned immutable real-prefix state"
117
+ },
118
+ "original_result_dir": "/root/dl/runtime/script/loopTF-workspace/output/benchmarks/20260920-000108_loop-full",
119
+ "original_policy": {
120
+ "num_passes": 3,
121
+ "depth_loops": 2,
122
+ "feedback_layers": [
123
+ 20
124
+ ],
125
+ "depth_share_kv": true,
126
+ "final_adapter": {
127
+ "kind": "lookahead_refiner",
128
+ "ffn_size": 5632,
129
+ "attention_backend": "sdpa_auto",
130
+ "checkpoint_refiner": true
131
+ }
132
+ },
133
+ "status": "succeeded",
134
+ "completed_tasks": 4,
135
+ "node": "22.169.130.39",
136
+ "gpu": 2,
137
+ "result_dir": "/root/dl/runtime/script/loopTF-workspace/output/benchmarks/20260920-140125_loop-prefix-full",
138
+ "tasks": {
139
+ "var_assign_d0_math_prefix_query": {
140
+ "loss": 0.9608591947823544,
141
+ "ppl": 2.613941393679959,
142
+ "accuracy": 0.686,
143
+ "loss_sum": 1564.2787691056728,
144
+ "num_tokens": 1628,
145
+ "examples": 1000,
146
+ "cropped_examples": 0
147
+ },
148
+ "var_assign_d0_code_prefix_query": {
149
+ "loss": 0.30558995497374086,
150
+ "ppl": 1.3574255868052258,
151
+ "accuracy": 0.886,
152
+ "loss_sum": 804.9239414008334,
153
+ "num_tokens": 2634,
154
+ "examples": 1000,
155
+ "cropped_examples": 0
156
+ },
157
+ "var_assign_d1_math_prefix_query": {
158
+ "loss": 2.3726828137036535,
159
+ "ppl": 10.726129926357867,
160
+ "accuracy": 0.066,
161
+ "loss_sum": 3924.417373865843,
162
+ "num_tokens": 1654,
163
+ "examples": 1000,
164
+ "cropped_examples": 0
165
+ },
166
+ "var_assign_d1_code_prefix_query": {
167
+ "loss": 1.072103428489692,
168
+ "ppl": 2.9215182462190192,
169
+ "accuracy": 0.301,
170
+ "loss_sum": 2849.650912925601,
171
+ "num_tokens": 2658,
172
+ "examples": 1000,
173
+ "cropped_examples": 0
174
+ }
175
+ },
176
+ "macro_loss": 1.17780884798736,
177
+ "macro_accuracy": 0.48475,
178
+ "original_tasks": {
179
+ "var_assign_d0_math": {
180
+ "loss": 0.5066267161589694,
181
+ "ppl": 1.659683159133497,
182
+ "accuracy": 0.589,
183
+ "loss_sum": 1331.4150100657716,
184
+ "num_tokens": 2628,
185
+ "examples": 1000,
186
+ "cropped_examples": 0
187
+ },
188
+ "var_assign_d0_code": {
189
+ "loss": 0.4721332660309784,
190
+ "ppl": 1.6034110497156713,
191
+ "accuracy": 0.653,
192
+ "loss_sum": 1243.599022725597,
193
+ "num_tokens": 2634,
194
+ "examples": 1000,
195
+ "cropped_examples": 0
196
+ },
197
+ "var_assign_d1_math": {
198
+ "loss": 0.8461934806654102,
199
+ "ppl": 2.330757869722473,
200
+ "accuracy": 0.246,
201
+ "loss_sum": 2245.7974976859987,
202
+ "num_tokens": 2654,
203
+ "examples": 1000,
204
+ "cropped_examples": 0
205
+ },
206
+ "var_assign_d1_code": {
207
+ "loss": 0.7977339243130903,
208
+ "ppl": 2.220503394176892,
209
+ "accuracy": 0.276,
210
+ "loss_sum": 2120.376770824194,
211
+ "num_tokens": 2658,
212
+ "examples": 1000,
213
+ "cropped_examples": 0
214
+ }
215
+ }
216
+ },
217
+ {
218
+ "name": "loop_adapter_26703",
219
+ "checkpoint_name": "paper2604_loop2_476m_7b_online_FBW0_K3_adapter_norm/step_26703",
220
+ "checkpoint": "/taijifs_wzbd31/rentTest/hunyuan/kenanchen/loopTF-workspace/output/paper2604-expanded-20260920/checkpoints/loop_adapter_26703/evaluation_native.pt",
221
+ "checkpoint_sha256": "3cdd26d737c0accf9e2c3098a3da2b0239ce7b9a577db51b6f309207555685e3",
222
+ "step": 26703,
223
+ "architecture": "full_bandwidth_loop",
224
+ "parameter_count": 653383808,
225
+ "training_tokens": 7000031232,
226
+ "inference": {
227
+ "requested": "auto",
228
+ "resolved": "fused",
229
+ "full_bandwidth": true,
230
+ "prompt_traversals": 2,
231
+ "fused_prefill_passes": 1,
232
+ "continuation": "one feedback transition per real token",
233
+ "cache": "architecture-owned immutable real-prefix state"
234
+ },
235
+ "original_result_dir": "/root/dl/runtime/script/loopTF-workspace/output/benchmarks/20260920-084128_loop_adapter_26703-full",
236
+ "original_policy": {
237
+ "num_passes": 3,
238
+ "depth_loops": 2,
239
+ "feedback_layers": [
240
+ 20
241
+ ],
242
+ "depth_share_kv": false,
243
+ "final_adapter": {
244
+ "kind": "lookahead_refiner",
245
+ "ffn_size": 5632,
246
+ "attention_backend": "sdpa_auto",
247
+ "checkpoint_refiner": true
248
+ }
249
+ },
250
+ "status": "succeeded",
251
+ "completed_tasks": 4,
252
+ "node": "22.169.130.39",
253
+ "gpu": 3,
254
+ "result_dir": "/root/dl/runtime/script/loopTF-workspace/output/benchmarks/20260920-140125_loop_adapter_26703-prefix-full",
255
+ "tasks": {
256
+ "var_assign_d0_math_prefix_query": {
257
+ "loss": 0.629065775921161,
258
+ "ppl": 1.875857289360853,
259
+ "accuracy": 0.814,
260
+ "loss_sum": 1024.11908319965,
261
+ "num_tokens": 1628,
262
+ "examples": 1000,
263
+ "cropped_examples": 0
264
+ },
265
+ "var_assign_d0_code_prefix_query": {
266
+ "loss": 0.20973520310653618,
267
+ "ppl": 1.2333514290862606,
268
+ "accuracy": 0.93,
269
+ "loss_sum": 552.4425249826163,
270
+ "num_tokens": 2634,
271
+ "examples": 1000,
272
+ "cropped_examples": 0
273
+ },
274
+ "var_assign_d1_math_prefix_query": {
275
+ "loss": 2.369307496793289,
276
+ "ppl": 10.689986870083283,
277
+ "accuracy": 0.036,
278
+ "loss_sum": 3918.8345996960998,
279
+ "num_tokens": 1654,
280
+ "examples": 1000,
281
+ "cropped_examples": 0
282
+ },
283
+ "var_assign_d1_code_prefix_query": {
284
+ "loss": 1.0343095231647101,
285
+ "ppl": 2.8131631408876365,
286
+ "accuracy": 0.291,
287
+ "loss_sum": 2749.1947125717998,
288
+ "num_tokens": 2658,
289
+ "examples": 1000,
290
+ "cropped_examples": 0
291
+ }
292
+ },
293
+ "macro_loss": 1.0606044997464241,
294
+ "macro_accuracy": 0.51775,
295
+ "original_tasks": {
296
+ "var_assign_d0_math": {
297
+ "loss": 0.22741154006851022,
298
+ "ppl": 1.2553463869869521,
299
+ "accuracy": 0.812,
300
+ "loss_sum": 597.6375273000449,
301
+ "num_tokens": 2628,
302
+ "examples": 1000,
303
+ "cropped_examples": 0
304
+ },
305
+ "var_assign_d0_code": {
306
+ "loss": 0.1440413290443466,
307
+ "ppl": 1.1549318397677866,
308
+ "accuracy": 0.943,
309
+ "loss_sum": 379.40486070280895,
310
+ "num_tokens": 2634,
311
+ "examples": 1000,
312
+ "cropped_examples": 0
313
+ },
314
+ "var_assign_d1_math": {
315
+ "loss": 0.7455220534322333,
316
+ "ppl": 2.107541397056164,
317
+ "accuracy": 0.292,
318
+ "loss_sum": 1978.6155298091471,
319
+ "num_tokens": 2654,
320
+ "examples": 1000,
321
+ "cropped_examples": 0
322
+ },
323
+ "var_assign_d1_code": {
324
+ "loss": 0.7551917844333534,
325
+ "ppl": 2.1280196052563283,
326
+ "accuracy": 0.302,
327
+ "loss_sum": 2007.2997630238533,
328
+ "num_tokens": 2658,
329
+ "examples": 1000,
330
+ "cropped_examples": 0
331
+ }
332
+ }
333
+ },
334
+ {
335
+ "name": "r3_d1_a2",
336
+ "checkpoint_name": "r3_d1_a2_r0_7b_seed42_20260919-035203-resume/step_26703",
337
+ "checkpoint": "/taijifs_wzbd31/rentTest/hunyuan/kenanchen/loopTF-workspace/output/paper2604-six-final-20260919/checkpoints/r3_d1_a2/evaluation_native.pt",
338
+ "checkpoint_sha256": "93bee6ad24da6497e3b5d41cbbd503ef7e387e1da7e48a11ab63dc639c8c6d0b",
339
+ "step": 26703,
340
+ "architecture": "lookahead",
341
+ "parameter_count": 659332608,
342
+ "training_tokens": 7000031232,
343
+ "inference": {
344
+ "requested": "auto",
345
+ "resolved": "fused",
346
+ "full_bandwidth": true,
347
+ "prompt_traversals": 2,
348
+ "fused_prefill_passes": 1,
349
+ "continuation": "one feedback transition per real token",
350
+ "cache": "architecture-owned immutable real-prefix state"
351
+ },
352
+ "original_result_dir": "/root/dl/runtime/script/loopTF-workspace/output/benchmarks/20260920-000113_r3_d1_a2-full",
353
+ "original_policy": {
354
+ "mode": "parallel",
355
+ "num_passes": 3,
356
+ "feedback_loss_weight": 1.0,
357
+ "attention_backend": "sdpa_auto",
358
+ "checkpoint_refiner": false,
359
+ "cache_layer_indices": null,
360
+ "refiner_intermediate_size": 5632,
361
+ "refiner_gate_bias": -2.0,
362
+ "future_loss_weight": 0.2,
363
+ "future_gradient_mode": "all",
364
+ "draft_loss_weight": 0.2,
365
+ "refined_future_loss_weight": 0.0,
366
+ "pseudo_initializer": "A2",
367
+ "pseudo_top_k": 128,
368
+ "pseudo_temperature": 1.0,
369
+ "refiner_variant": "R0",
370
+ "allow_objective_change_on_resume": false
371
+ },
372
+ "status": "succeeded",
373
+ "completed_tasks": 4,
374
+ "node": "22.169.167.168",
375
+ "gpu": 0,
376
+ "result_dir": "/root/dl/runtime/script/loopTF-workspace/output/benchmarks/20260920-140027_r3_d1_a2-prefix-full",
377
+ "tasks": {
378
+ "var_assign_d0_math_prefix_query": {
379
+ "loss": 0.9182237858039668,
380
+ "ppl": 2.504837308707536,
381
+ "accuracy": 0.65,
382
+ "loss_sum": 1494.868323288858,
383
+ "num_tokens": 1628,
384
+ "examples": 1000,
385
+ "cropped_examples": 0
386
+ },
387
+ "var_assign_d0_code_prefix_query": {
388
+ "loss": 0.28392734452310403,
389
+ "ppl": 1.3283364163308797,
390
+ "accuracy": 0.914,
391
+ "loss_sum": 747.864625473856,
392
+ "num_tokens": 2634,
393
+ "examples": 1000,
394
+ "cropped_examples": 0
395
+ },
396
+ "var_assign_d1_math_prefix_query": {
397
+ "loss": 2.633833445941955,
398
+ "ppl": 13.927056319105292,
399
+ "accuracy": 0.035,
400
+ "loss_sum": 4356.360519587994,
401
+ "num_tokens": 1654,
402
+ "examples": 1000,
403
+ "cropped_examples": 0
404
+ },
405
+ "var_assign_d1_code_prefix_query": {
406
+ "loss": 1.286979318492008,
407
+ "ppl": 3.6218296216169215,
408
+ "accuracy": 0.156,
409
+ "loss_sum": 3420.7910285517573,
410
+ "num_tokens": 2658,
411
+ "examples": 1000,
412
+ "cropped_examples": 0
413
+ }
414
+ },
415
+ "macro_loss": 1.2807409736902584,
416
+ "macro_accuracy": 0.43875000000000003,
417
+ "original_tasks": {
418
+ "var_assign_d0_math": {
419
+ "loss": 0.3188153027884288,
420
+ "ppl": 1.375497250936613,
421
+ "accuracy": 0.762,
422
+ "loss_sum": 837.8466157279909,
423
+ "num_tokens": 2628,
424
+ "examples": 1000,
425
+ "cropped_examples": 0
426
+ },
427
+ "var_assign_d0_code": {
428
+ "loss": 0.32760351651858033,
429
+ "ppl": 1.387638687389456,
430
+ "accuracy": 0.793,
431
+ "loss_sum": 862.9076625099406,
432
+ "num_tokens": 2634,
433
+ "examples": 1000,
434
+ "cropped_examples": 0
435
+ },
436
+ "var_assign_d1_math": {
437
+ "loss": 0.8227609494089035,
438
+ "ppl": 2.2767772345168704,
439
+ "accuracy": 0.256,
440
+ "loss_sum": 2183.60755973123,
441
+ "num_tokens": 2654,
442
+ "examples": 1000,
443
+ "cropped_examples": 0
444
+ },
445
+ "var_assign_d1_code": {
446
+ "loss": 0.827892017831775,
447
+ "ppl": 2.28848955695804,
448
+ "accuracy": 0.261,
449
+ "loss_sum": 2200.536983396858,
450
+ "num_tokens": 2658,
451
+ "examples": 1000,
452
+ "cropped_examples": 0
453
+ }
454
+ }
455
+ },
456
+ {
457
+ "name": "alpha0_seed43_26703",
458
+ "checkpoint_name": "r3_b300_alpha0p0-seed43_product_20260918-170109/step_26703",
459
+ "checkpoint": "/taijifs_wzbd31/rentTest/hunyuan/kenanchen/loopTF-workspace/output/paper2604-expanded-20260920/checkpoints/alpha0_seed43_26703/evaluation_native.pt",
460
+ "checkpoint_sha256": "29330017a0a088deeb43f224e0cb3d7f7608d253807ff32d8f2addfd6a4edee4",
461
+ "step": 26703,
462
+ "architecture": "lookahead",
463
+ "parameter_count": 661315072,
464
+ "training_tokens": 7000031232,
465
+ "inference": {
466
+ "requested": "auto",
467
+ "resolved": "fused",
468
+ "full_bandwidth": true,
469
+ "prompt_traversals": 2,
470
+ "fused_prefill_passes": 1,
471
+ "continuation": "one feedback transition per real token",
472
+ "cache": "architecture-owned immutable real-prefix state"
473
+ },
474
+ "original_result_dir": "/root/dl/runtime/script/loopTF-workspace/output/benchmarks/20260920-103100_alpha0_seed43_26703-full",
475
+ "original_policy": {
476
+ "mode": "parallel",
477
+ "num_passes": 3,
478
+ "feedback_loss_weight": 1.0,
479
+ "attention_backend": "sdpa_auto",
480
+ "checkpoint_refiner": false,
481
+ "cache_layer_indices": null,
482
+ "refiner_intermediate_size": 5632,
483
+ "refiner_gate_bias": -2.0,
484
+ "future_loss_weight": 0.0,
485
+ "future_gradient_mode": "all",
486
+ "draft_loss_weight": 0.0,
487
+ "refined_future_loss_weight": 0.0,
488
+ "pseudo_initializer": "A0",
489
+ "pseudo_top_k": 128,
490
+ "pseudo_temperature": 1.0,
491
+ "refiner_variant": "R0",
492
+ "allow_objective_change_on_resume": false
493
+ },
494
+ "status": "succeeded",
495
+ "completed_tasks": 4,
496
+ "node": "22.169.167.168",
497
+ "gpu": 1,
498
+ "result_dir": "/root/dl/runtime/script/loopTF-workspace/output/benchmarks/20260920-140027_alpha0_seed43_26703-prefix-full",
499
+ "tasks": {
500
+ "var_assign_d0_math_prefix_query": {
501
+ "loss": 0.8356628586567884,
502
+ "ppl": 2.306342320689545,
503
+ "accuracy": 0.764,
504
+ "loss_sum": 1360.4591338932514,
505
+ "num_tokens": 1628,
506
+ "examples": 1000,
507
+ "cropped_examples": 0
508
+ },
509
+ "var_assign_d0_code_prefix_query": {
510
+ "loss": 0.5161943806986973,
511
+ "ppl": 1.6756386575566136,
512
+ "accuracy": 0.759,
513
+ "loss_sum": 1359.6559987603687,
514
+ "num_tokens": 2634,
515
+ "examples": 1000,
516
+ "cropped_examples": 0
517
+ },
518
+ "var_assign_d1_math_prefix_query": {
519
+ "loss": 2.794346192493548,
520
+ "ppl": 16.35193424011053,
521
+ "accuracy": 0.021,
522
+ "loss_sum": 4621.848602384329,
523
+ "num_tokens": 1654,
524
+ "examples": 1000,
525
+ "cropped_examples": 0
526
+ },
527
+ "var_assign_d1_code_prefix_query": {
528
+ "loss": 1.4300516922954691,
529
+ "ppl": 4.178915204059585,
530
+ "accuracy": 0.157,
531
+ "loss_sum": 3801.077398121357,
532
+ "num_tokens": 2658,
533
+ "examples": 1000,
534
+ "cropped_examples": 0
535
+ }
536
+ },
537
+ "macro_loss": 1.3940637810361258,
538
+ "macro_accuracy": 0.42525,
539
+ "original_tasks": {
540
+ "var_assign_d0_math": {
541
+ "loss": 0.2549581390091812,
542
+ "ppl": 1.290407602001476,
543
+ "accuracy": 0.851,
544
+ "loss_sum": 670.0299893161282,
545
+ "num_tokens": 2628,
546
+ "examples": 1000,
547
+ "cropped_examples": 0
548
+ },
549
+ "var_assign_d0_code": {
550
+ "loss": 0.2881996440644614,
551
+ "ppl": 1.3340236074346379,
552
+ "accuracy": 0.859,
553
+ "loss_sum": 759.1178624657914,
554
+ "num_tokens": 2634,
555
+ "examples": 1000,
556
+ "cropped_examples": 0
557
+ },
558
+ "var_assign_d1_math": {
559
+ "loss": 0.8654673991008285,
560
+ "ppl": 2.376116421150777,
561
+ "accuracy": 0.228,
562
+ "loss_sum": 2296.950477213599,
563
+ "num_tokens": 2654,
564
+ "examples": 1000,
565
+ "cropped_examples": 0
566
+ },
567
+ "var_assign_d1_code": {
568
+ "loss": 0.8612851583461247,
569
+ "ppl": 2.3661996817667057,
570
+ "accuracy": 0.267,
571
+ "loss_sum": 2289.2959508839995,
572
+ "num_tokens": 2658,
573
+ "examples": 1000,
574
+ "cropped_examples": 0
575
+ }
576
+ }
577
+ },
578
+ {
579
+ "name": "fbw_l2_l11_l20_26703",
580
+ "checkpoint_name": "fbw_L2L11L20_7b_seed42_20260920-023045/step_26703",
581
+ "checkpoint": "/taijifs_wzbd31/rentTest/hunyuan/kenanchen/loopTF-workspace/output/paper2604-all-20260920/checkpoints/fbw_l2_l11_l20_26703/evaluation_native.pt",
582
+ "checkpoint_sha256": "8a4995528738017ec28df7d00ded356dc6c1b35688b14972a35b130d4e2932d4",
583
+ "step": 26703,
584
+ "architecture": "full_bandwidth",
585
+ "parameter_count": 576059264,
586
+ "training_tokens": 7000031232,
587
+ "inference": {
588
+ "requested": "auto",
589
+ "resolved": "fused",
590
+ "full_bandwidth": true,
591
+ "prompt_traversals": 2,
592
+ "fused_prefill_passes": 1,
593
+ "continuation": "one feedback transition per real token",
594
+ "cache": "architecture-owned immutable real-prefix state"
595
+ },
596
+ "original_result_dir": "/root/dl/runtime/script/loopTF-workspace/output/benchmarks/20260920-103238_fbw_l2_l11_l20_26703-full",
597
+ "original_policy": {
598
+ "num_passes": 3,
599
+ "depth_loops": 1,
600
+ "feedback_layers": [
601
+ 2,
602
+ 11,
603
+ 20
604
+ ],
605
+ "depth_share_kv": false,
606
+ "final_adapter": {
607
+ "kind": "none",
608
+ "ffn_size": null,
609
+ "attention_backend": "sdpa_auto",
610
+ "checkpoint_refiner": true
611
+ }
612
+ },
613
+ "status": "succeeded",
614
+ "completed_tasks": 4,
615
+ "node": "22.169.167.168",
616
+ "gpu": 4,
617
+ "result_dir": "/root/dl/runtime/script/loopTF-workspace/output/benchmarks/20260920-140028_fbw_l2_l11_l20_26703-prefix-full",
618
+ "tasks": {
619
+ "var_assign_d0_math_prefix_query": {
620
+ "loss": 0.6999826131603641,
621
+ "ppl": 2.0137176949794644,
622
+ "accuracy": 0.899,
623
+ "loss_sum": 1139.5716942250729,
624
+ "num_tokens": 1628,
625
+ "examples": 1000,
626
+ "cropped_examples": 0
627
+ },
628
+ "var_assign_d0_code_prefix_query": {
629
+ "loss": 0.1681138649393699,
630
+ "ppl": 1.1830713133220712,
631
+ "accuracy": 0.966,
632
+ "loss_sum": 442.8119202503003,
633
+ "num_tokens": 2634,
634
+ "examples": 1000,
635
+ "cropped_examples": 0
636
+ },
637
+ "var_assign_d1_math_prefix_query": {
638
+ "loss": 2.289529290343066,
639
+ "ppl": 9.870290546456431,
640
+ "accuracy": 0.085,
641
+ "loss_sum": 3786.8814462274313,
642
+ "num_tokens": 1654,
643
+ "examples": 1000,
644
+ "cropped_examples": 0
645
+ },
646
+ "var_assign_d1_code_prefix_query": {
647
+ "loss": 1.1037676931927198,
648
+ "ppl": 3.0155061494666158,
649
+ "accuracy": 0.275,
650
+ "loss_sum": 2933.814528506249,
651
+ "num_tokens": 2658,
652
+ "examples": 1000,
653
+ "cropped_examples": 0
654
+ }
655
+ },
656
+ "macro_loss": 1.06534836540888,
657
+ "macro_accuracy": 0.55625,
658
+ "original_tasks": {
659
+ "var_assign_d0_math": {
660
+ "loss": 0.5708084764764212,
661
+ "ppl": 1.769697231804542,
662
+ "accuracy": 0.451,
663
+ "loss_sum": 1500.0846761800349,
664
+ "num_tokens": 2628,
665
+ "examples": 1000,
666
+ "cropped_examples": 0
667
+ },
668
+ "var_assign_d0_code": {
669
+ "loss": 0.5145805300713355,
670
+ "ppl": 1.6729366079965335,
671
+ "accuracy": 0.533,
672
+ "loss_sum": 1355.4051162078977,
673
+ "num_tokens": 2634,
674
+ "examples": 1000,
675
+ "cropped_examples": 0
676
+ },
677
+ "var_assign_d1_math": {
678
+ "loss": 0.8756185885843033,
679
+ "ppl": 2.4003596699007836,
680
+ "accuracy": 0.223,
681
+ "loss_sum": 2323.891734102741,
682
+ "num_tokens": 2654,
683
+ "examples": 1000,
684
+ "cropped_examples": 0
685
+ },
686
+ "var_assign_d1_code": {
687
+ "loss": 0.8256150507841167,
688
+ "ppl": 2.2832846695752274,
689
+ "accuracy": 0.263,
690
+ "loss_sum": 2194.484804984182,
691
+ "num_tokens": 2658,
692
+ "examples": 1000,
693
+ "cropped_examples": 0
694
+ }
695
+ }
696
+ },
697
+ {
698
+ "name": "lookahead",
699
+ "checkpoint_name": "paper2604_lookahead_476m_7b_online_K3_LA2_20260917-190928/step_26703",
700
+ "checkpoint": "/taijifs_wzbd31/rentTest/hunyuan/kenanchen/loopTF-workspace/output/evaluation-readiness/checkpoints/lookahead/evaluation_native.pt",
701
+ "checkpoint_sha256": "1371d5d6d86d0776fffa4fe9e7b7da5592acb01ce2b60e240a459f86028bff82",
702
+ "step": 26703,
703
+ "architecture": "lookahead",
704
+ "parameter_count": 661315072,
705
+ "training_tokens": 7000031232,
706
+ "inference": {
707
+ "requested": "auto",
708
+ "resolved": "fused",
709
+ "full_bandwidth": true,
710
+ "prompt_traversals": 2,
711
+ "fused_prefill_passes": 1,
712
+ "continuation": "one feedback transition per real token",
713
+ "cache": "architecture-owned immutable real-prefix state"
714
+ },
715
+ "original_result_dir": "/root/dl/runtime/script/loopTF-workspace/output/benchmarks/20260920-000107_lookahead-full",
716
+ "original_policy": {
717
+ "mode": "parallel",
718
+ "num_passes": 3,
719
+ "feedback_loss_weight": 1.0,
720
+ "attention_backend": "sdpa_auto",
721
+ "checkpoint_refiner": false,
722
+ "cache_layer_indices": null,
723
+ "refiner_intermediate_size": 5632,
724
+ "refiner_gate_bias": -2.0,
725
+ "future_loss_weight": 0.0,
726
+ "future_gradient_mode": "all",
727
+ "draft_loss_weight": 0.0,
728
+ "refined_future_loss_weight": 0.0,
729
+ "pseudo_initializer": "A0",
730
+ "pseudo_top_k": 128,
731
+ "pseudo_temperature": 1.0,
732
+ "refiner_variant": "R0",
733
+ "allow_objective_change_on_resume": false
734
+ },
735
+ "status": "succeeded",
736
+ "completed_tasks": 4,
737
+ "node": "22.169.167.168",
738
+ "gpu": 5,
739
+ "result_dir": "/root/dl/runtime/script/loopTF-workspace/output/benchmarks/20260920-140030_lookahead-prefix-full",
740
+ "tasks": {
741
+ "var_assign_d0_math_prefix_query": {
742
+ "loss": 0.9661379674362609,
743
+ "ppl": 2.627776279510075,
744
+ "accuracy": 0.732,
745
+ "loss_sum": 1572.8726109862328,
746
+ "num_tokens": 1628,
747
+ "examples": 1000,
748
+ "cropped_examples": 0
749
+ },
750
+ "var_assign_d0_code_prefix_query": {
751
+ "loss": 0.33066893643622686,
752
+ "ppl": 1.3918989090083416,
753
+ "accuracy": 0.908,
754
+ "loss_sum": 870.9819785730215,
755
+ "num_tokens": 2634,
756
+ "examples": 1000,
757
+ "cropped_examples": 0
758
+ },
759
+ "var_assign_d1_math_prefix_query": {
760
+ "loss": 2.242093697208303,
761
+ "ppl": 9.413018681465168,
762
+ "accuracy": 0.098,
763
+ "loss_sum": 3708.4229751825333,
764
+ "num_tokens": 1654,
765
+ "examples": 1000,
766
+ "cropped_examples": 0
767
+ },
768
+ "var_assign_d1_code_prefix_query": {
769
+ "loss": 1.3947670557406542,
770
+ "ppl": 4.034034757935371,
771
+ "accuracy": 0.139,
772
+ "loss_sum": 3707.290834158659,
773
+ "num_tokens": 2658,
774
+ "examples": 1000,
775
+ "cropped_examples": 0
776
+ }
777
+ },
778
+ "macro_loss": 1.2334169142053613,
779
+ "macro_accuracy": 0.46925,
780
+ "original_tasks": {
781
+ "var_assign_d0_math": {
782
+ "loss": 0.49276460260379623,
783
+ "ppl": 1.6368351692986427,
784
+ "accuracy": 0.582,
785
+ "loss_sum": 1294.9853756427765,
786
+ "num_tokens": 2628,
787
+ "examples": 1000,
788
+ "cropped_examples": 0
789
+ },
790
+ "var_assign_d0_code": {
791
+ "loss": 0.3921574250064456,
792
+ "ppl": 1.4801707089453837,
793
+ "accuracy": 0.783,
794
+ "loss_sum": 1032.9426574669778,
795
+ "num_tokens": 2634,
796
+ "examples": 1000,
797
+ "cropped_examples": 0
798
+ },
799
+ "var_assign_d1_math": {
800
+ "loss": 0.7952725603003069,
801
+ "ppl": 2.215044647769082,
802
+ "accuracy": 0.26,
803
+ "loss_sum": 2110.6533750370145,
804
+ "num_tokens": 2654,
805
+ "examples": 1000,
806
+ "cropped_examples": 0
807
+ },
808
+ "var_assign_d1_code": {
809
+ "loss": 0.8039420324442887,
810
+ "ppl": 2.234331397812932,
811
+ "accuracy": 0.291,
812
+ "loss_sum": 2136.8779222369194,
813
+ "num_tokens": 2658,
814
+ "examples": 1000,
815
+ "cropped_examples": 0
816
+ }
817
+ }
818
+ },
819
+ {
820
+ "name": "r3_d0_a2",
821
+ "checkpoint_name": "r3_D0-A2-R0_7b_seed42_20260919-040210/step_26703",
822
+ "checkpoint": "/taijifs_wzbd31/rentTest/hunyuan/kenanchen/loopTF-workspace/output/paper2604-six-final-20260919/checkpoints/r3_d0_a2/evaluation_native.pt",
823
+ "checkpoint_sha256": "3b7b18135b7849c0db3b4a1f8dd737911616b35993af54f8ee3f8eba8e254274",
824
+ "step": 26703,
825
+ "architecture": "lookahead",
826
+ "parameter_count": 659332608,
827
+ "training_tokens": 7000031232,
828
+ "inference": {
829
+ "requested": "auto",
830
+ "resolved": "fused",
831
+ "full_bandwidth": true,
832
+ "prompt_traversals": 2,
833
+ "fused_prefill_passes": 1,
834
+ "continuation": "one feedback transition per real token",
835
+ "cache": "architecture-owned immutable real-prefix state"
836
+ },
837
+ "original_result_dir": "/root/dl/runtime/script/loopTF-workspace/output/benchmarks/20260920-000110_r3_d0_a2-full",
838
+ "original_policy": {
839
+ "mode": "parallel",
840
+ "num_passes": 3,
841
+ "feedback_loss_weight": 1.0,
842
+ "attention_backend": "sdpa_auto",
843
+ "checkpoint_refiner": false,
844
+ "cache_layer_indices": null,
845
+ "refiner_intermediate_size": 5632,
846
+ "refiner_gate_bias": -2.0,
847
+ "future_loss_weight": 0.2,
848
+ "future_gradient_mode": "all",
849
+ "draft_loss_weight": 0.0,
850
+ "refined_future_loss_weight": 0.0,
851
+ "pseudo_initializer": "A2",
852
+ "pseudo_top_k": 128,
853
+ "pseudo_temperature": 1.0,
854
+ "refiner_variant": "R0",
855
+ "allow_objective_change_on_resume": false
856
+ },
857
+ "status": "succeeded",
858
+ "completed_tasks": 4,
859
+ "node": "22.169.167.168",
860
+ "gpu": 6,
861
+ "result_dir": "/root/dl/runtime/script/loopTF-workspace/output/benchmarks/20260920-140033_r3_d0_a2-prefix-full",
862
+ "tasks": {
863
+ "var_assign_d0_math_prefix_query": {
864
+ "loss": 1.211432274211317,
865
+ "ppl": 3.3582912034626933,
866
+ "accuracy": 0.58,
867
+ "loss_sum": 1972.2117424160242,
868
+ "num_tokens": 1628,
869
+ "examples": 1000,
870
+ "cropped_examples": 0
871
+ },
872
+ "var_assign_d0_code_prefix_query": {
873
+ "loss": 0.37923464322343875,
874
+ "ppl": 1.461165848187148,
875
+ "accuracy": 0.836,
876
+ "loss_sum": 998.9040502505377,
877
+ "num_tokens": 2634,
878
+ "examples": 1000,
879
+ "cropped_examples": 0
880
+ },
881
+ "var_assign_d1_math_prefix_query": {
882
+ "loss": 2.7344340565631295,
883
+ "ppl": 15.401024860110988,
884
+ "accuracy": 0.026,
885
+ "loss_sum": 4522.753929555416,
886
+ "num_tokens": 1654,
887
+ "examples": 1000,
888
+ "cropped_examples": 0
889
+ },
890
+ "var_assign_d1_code_prefix_query": {
891
+ "loss": 1.4339815225813781,
892
+ "ppl": 4.195369942578986,
893
+ "accuracy": 0.203,
894
+ "loss_sum": 3811.522887021303,
895
+ "num_tokens": 2658,
896
+ "examples": 1000,
897
+ "cropped_examples": 0
898
+ }
899
+ },
900
+ "macro_loss": 1.4397706241448158,
901
+ "macro_accuracy": 0.41125,
902
+ "original_tasks": {
903
+ "var_assign_d0_math": {
904
+ "loss": 0.5558424301869328,
905
+ "ppl": 1.7434090667349615,
906
+ "accuracy": 0.512,
907
+ "loss_sum": 1460.7539065312594,
908
+ "num_tokens": 2628,
909
+ "examples": 1000,
910
+ "cropped_examples": 0
911
+ },
912
+ "var_assign_d0_code": {
913
+ "loss": 0.4339058975757707,
914
+ "ppl": 1.5432736354928742,
915
+ "accuracy": 0.772,
916
+ "loss_sum": 1142.90813421458,
917
+ "num_tokens": 2634,
918
+ "examples": 1000,
919
+ "cropped_examples": 0
920
+ },
921
+ "var_assign_d1_math": {
922
+ "loss": 0.8644822358119371,
923
+ "ppl": 2.37377671116995,
924
+ "accuracy": 0.247,
925
+ "loss_sum": 2294.335853844881,
926
+ "num_tokens": 2654,
927
+ "examples": 1000,
928
+ "cropped_examples": 0
929
+ },
930
+ "var_assign_d1_code": {
931
+ "loss": 0.8591235937821161,
932
+ "ppl": 2.3610905122724195,
933
+ "accuracy": 0.276,
934
+ "loss_sum": 2283.5505122728646,
935
+ "num_tokens": 2658,
936
+ "examples": 1000,
937
+ "cropped_examples": 0
938
+ }
939
+ }
940
+ },
941
+ {
942
+ "name": "r3_d0_p0_a0_r1_26703",
943
+ "checkpoint_name": "r3_D0-P0-A0-R1_7b_seed42_20260919-161651/step_26703",
944
+ "checkpoint": "/taijifs_wzbd31/rentTest/hunyuan/kenanchen/loopTF-workspace/output/paper2604-all-20260920/checkpoints/r3_d0_p0_a0_r1_26703/evaluation_native.pt",
945
+ "checkpoint_sha256": "834dfd0ce5e0214673c3a976ce08ba7694bf77995b0b89e4b7d21faa39af9021",
946
+ "step": 26703,
947
+ "architecture": "lookahead",
948
+ "parameter_count": 667258240,
949
+ "training_tokens": 7000031232,
950
+ "inference": {
951
+ "requested": "auto",
952
+ "resolved": "fused",
953
+ "full_bandwidth": true,
954
+ "prompt_traversals": 2,
955
+ "fused_prefill_passes": 1,
956
+ "continuation": "one feedback transition per real token",
957
+ "cache": "architecture-owned immutable real-prefix state"
958
+ },
959
+ "original_result_dir": "/root/dl/runtime/script/loopTF-workspace/output/benchmarks/20260920-103500_r3_d0_p0_a0_r1_26703-full",
960
+ "original_policy": {
961
+ "mode": "parallel",
962
+ "num_passes": 3,
963
+ "feedback_loss_weight": 1.0,
964
+ "attention_backend": "sdpa_auto",
965
+ "checkpoint_refiner": false,
966
+ "cache_layer_indices": null,
967
+ "refiner_intermediate_size": 5632,
968
+ "refiner_gate_bias": -2.0,
969
+ "future_loss_weight": 0.0,
970
+ "future_gradient_mode": "all",
971
+ "draft_loss_weight": 0.0,
972
+ "refined_future_loss_weight": 1.0,
973
+ "pseudo_initializer": "A0",
974
+ "pseudo_top_k": 128,
975
+ "pseudo_temperature": 1.0,
976
+ "refiner_variant": "R1",
977
+ "allow_objective_change_on_resume": false
978
+ },
979
+ "status": "succeeded",
980
+ "completed_tasks": 4,
981
+ "node": "22.169.167.168",
982
+ "gpu": 7,
983
+ "result_dir": "/root/dl/runtime/script/loopTF-workspace/output/benchmarks/20260920-140033_r3_d0_p0_a0_r1_26703-prefix-full",
984
+ "tasks": {
985
+ "var_assign_d0_math_prefix_query": {
986
+ "loss": 0.7044368931080881,
987
+ "ppl": 2.022707363715812,
988
+ "accuracy": 0.825,
989
+ "loss_sum": 1146.8232619799674,
990
+ "num_tokens": 1628,
991
+ "examples": 1000,
992
+ "cropped_examples": 0
993
+ },
994
+ "var_assign_d0_code_prefix_query": {
995
+ "loss": 0.27659975864576924,
996
+ "ppl": 1.3186384917467677,
997
+ "accuracy": 0.897,
998
+ "loss_sum": 728.5637642729562,
999
+ "num_tokens": 2634,
1000
+ "examples": 1000,
1001
+ "cropped_examples": 0
1002
+ },
1003
+ "var_assign_d1_math_prefix_query": {
1004
+ "loss": 1.887181619382683,
1005
+ "ppl": 6.60073904568683,
1006
+ "accuracy": 0.165,
1007
+ "loss_sum": 3121.3983984589577,
1008
+ "num_tokens": 1654,
1009
+ "examples": 1000,
1010
+ "cropped_examples": 0
1011
+ },
1012
+ "var_assign_d1_code_prefix_query": {
1013
+ "loss": 1.1083835366695693,
1014
+ "ppl": 3.029457427543485,
1015
+ "accuracy": 0.248,
1016
+ "loss_sum": 2946.0834404677153,
1017
+ "num_tokens": 2658,
1018
+ "examples": 1000,
1019
+ "cropped_examples": 0
1020
+ }
1021
+ },
1022
+ "macro_loss": 0.9941504519515274,
1023
+ "macro_accuracy": 0.53375,
1024
+ "original_tasks": {
1025
+ "var_assign_d0_math": {
1026
+ "loss": 0.4598649001408856,
1027
+ "ppl": 1.5838599912778681,
1028
+ "accuracy": 0.491,
1029
+ "loss_sum": 1208.5249575702474,
1030
+ "num_tokens": 2628,
1031
+ "examples": 1000,
1032
+ "cropped_examples": 0
1033
+ },
1034
+ "var_assign_d0_code": {
1035
+ "loss": 0.37097813815621683,
1036
+ "ppl": 1.4491513919189898,
1037
+ "accuracy": 0.657,
1038
+ "loss_sum": 977.1564159034751,
1039
+ "num_tokens": 2634,
1040
+ "examples": 1000,
1041
+ "cropped_examples": 0
1042
+ },
1043
+ "var_assign_d1_math": {
1044
+ "loss": 0.734794200127825,
1045
+ "ppl": 2.0850528447383394,
1046
+ "accuracy": 0.259,
1047
+ "loss_sum": 1950.1438071392477,
1048
+ "num_tokens": 2654,
1049
+ "examples": 1000,
1050
+ "cropped_examples": 0
1051
+ },
1052
+ "var_assign_d1_code": {
1053
+ "loss": 0.7708381090840175,
1054
+ "ppl": 2.1615771322524124,
1055
+ "accuracy": 0.284,
1056
+ "loss_sum": 2048.8876939453185,
1057
+ "num_tokens": 2658,
1058
+ "examples": 1000,
1059
+ "cropped_examples": 0
1060
+ }
1061
+ }
1062
+ },
1063
+ {
1064
+ "name": "r3_d0_p0_a2_26703",
1065
+ "checkpoint_name": "r3_D0-P0-A2-R0_7b_seed42_20260919-160042/step_26703",
1066
+ "checkpoint": "/taijifs_wzbd31/rentTest/hunyuan/kenanchen/loopTF-workspace/output/paper2604-expanded-20260920/checkpoints/r3_d0_p0_a2_26703/evaluation_native.pt",
1067
+ "checkpoint_sha256": "c780e1e5a3057cb50d25b26d07a2ce4d665a8b5808c5070670e18921c5d12073",
1068
+ "step": 26703,
1069
+ "architecture": "lookahead",
1070
+ "parameter_count": 659332608,
1071
+ "training_tokens": 7000031232,
1072
+ "inference": {
1073
+ "requested": "auto",
1074
+ "resolved": "fused",
1075
+ "full_bandwidth": true,
1076
+ "prompt_traversals": 2,
1077
+ "fused_prefill_passes": 1,
1078
+ "continuation": "one feedback transition per real token",
1079
+ "cache": "architecture-owned immutable real-prefix state"
1080
+ },
1081
+ "original_result_dir": "/root/dl/runtime/script/loopTF-workspace/output/benchmarks/20260920-084131_r3_d0_p0_a2_26703-full",
1082
+ "original_policy": {
1083
+ "mode": "parallel",
1084
+ "num_passes": 3,
1085
+ "feedback_loss_weight": 1.0,
1086
+ "attention_backend": "sdpa_auto",
1087
+ "checkpoint_refiner": false,
1088
+ "cache_layer_indices": null,
1089
+ "refiner_intermediate_size": 5632,
1090
+ "refiner_gate_bias": -2.0,
1091
+ "future_loss_weight": 0.0,
1092
+ "future_gradient_mode": "all",
1093
+ "draft_loss_weight": 0.0,
1094
+ "refined_future_loss_weight": 0.0,
1095
+ "pseudo_initializer": "A2",
1096
+ "pseudo_top_k": 128,
1097
+ "pseudo_temperature": 1.0,
1098
+ "refiner_variant": "R0",
1099
+ "allow_objective_change_on_resume": false
1100
+ },
1101
+ "status": "succeeded",
1102
+ "completed_tasks": 4,
1103
+ "node": "22.169.167.168",
1104
+ "gpu": 4,
1105
+ "result_dir": "/root/dl/runtime/script/loopTF-workspace/output/benchmarks/20260920-140431_r3_d0_p0_a2_26703-prefix-full",
1106
+ "tasks": {
1107
+ "var_assign_d0_math_prefix_query": {
1108
+ "loss": 0.5820105113340793,
1109
+ "ppl": 1.7896328934013663,
1110
+ "accuracy": 0.902,
1111
+ "loss_sum": 947.5131124518812,
1112
+ "num_tokens": 1628,
1113
+ "examples": 1000,
1114
+ "cropped_examples": 0
1115
+ },
1116
+ "var_assign_d0_code_prefix_query": {
1117
+ "loss": 0.2823917204030629,
1118
+ "ppl": 1.3262981562916147,
1119
+ "accuracy": 0.964,
1120
+ "loss_sum": 743.8197915416677,
1121
+ "num_tokens": 2634,
1122
+ "examples": 1000,
1123
+ "cropped_examples": 0
1124
+ },
1125
+ "var_assign_d1_math_prefix_query": {
1126
+ "loss": 2.3664213439681654,
1127
+ "ppl": 10.659178414618935,
1128
+ "accuracy": 0.034,
1129
+ "loss_sum": 3914.0609029233456,
1130
+ "num_tokens": 1654,
1131
+ "examples": 1000,
1132
+ "cropped_examples": 0
1133
+ },
1134
+ "var_assign_d1_code_prefix_query": {
1135
+ "loss": 1.2515004484432815,
1136
+ "ppl": 3.4955839680687095,
1137
+ "accuracy": 0.195,
1138
+ "loss_sum": 3326.488191962242,
1139
+ "num_tokens": 2658,
1140
+ "examples": 1000,
1141
+ "cropped_examples": 0
1142
+ }
1143
+ },
1144
+ "macro_loss": 1.1205810060371473,
1145
+ "macro_accuracy": 0.52375,
1146
+ "original_tasks": {
1147
+ "var_assign_d0_math": {
1148
+ "loss": 0.3377159218682342,
1149
+ "ppl": 1.4017422425219221,
1150
+ "accuracy": 0.842,
1151
+ "loss_sum": 887.5174426697195,
1152
+ "num_tokens": 2628,
1153
+ "examples": 1000,
1154
+ "cropped_examples": 0
1155
+ },
1156
+ "var_assign_d0_code": {
1157
+ "loss": 0.47410123599671966,
1158
+ "ppl": 1.6065696214720206,
1159
+ "accuracy": 0.808,
1160
+ "loss_sum": 1248.7826556153595,
1161
+ "num_tokens": 2634,
1162
+ "examples": 1000,
1163
+ "cropped_examples": 0
1164
+ },
1165
+ "var_assign_d1_math": {
1166
+ "loss": 0.7743165313400266,
1167
+ "ppl": 2.169109102349156,
1168
+ "accuracy": 0.257,
1169
+ "loss_sum": 2055.0360741764307,
1170
+ "num_tokens": 2654,
1171
+ "examples": 1000,
1172
+ "cropped_examples": 0
1173
+ },
1174
+ "var_assign_d1_code": {
1175
+ "loss": 0.9058389874368603,
1176
+ "ppl": 2.474006713125388,
1177
+ "accuracy": 0.266,
1178
+ "loss_sum": 2407.7200286071748,
1179
+ "num_tokens": 2658,
1180
+ "examples": 1000,
1181
+ "cropped_examples": 0
1182
+ }
1183
+ }
1184
+ },
1185
+ {
1186
+ "name": "r3_d0_p0_a2_r1_26703",
1187
+ "checkpoint_name": "r3_D0-P0-A2-R1_7b_seed42_20260919-161651/step_26703",
1188
+ "checkpoint": "/taijifs_wzbd31/rentTest/hunyuan/kenanchen/loopTF-workspace/output/paper2604-all-20260920/checkpoints/r3_d0_p0_a2_r1_26703/evaluation_native.pt",
1189
+ "checkpoint_sha256": "abd0d7bda64f75c1a9eecdaa40f8aa7aaad176e1873358d495cd181bd29c3764",
1190
+ "step": 26703,
1191
+ "architecture": "lookahead",
1192
+ "parameter_count": 665275776,
1193
+ "training_tokens": 7000031232,
1194
+ "inference": {
1195
+ "requested": "auto",
1196
+ "resolved": "fused",
1197
+ "full_bandwidth": true,
1198
+ "prompt_traversals": 2,
1199
+ "fused_prefill_passes": 1,
1200
+ "continuation": "one feedback transition per real token",
1201
+ "cache": "architecture-owned immutable real-prefix state"
1202
+ },
1203
+ "original_result_dir": "/root/dl/runtime/script/loopTF-workspace/output/benchmarks/20260920-114704_r3_d0_p0_a2_r1_26703-full",
1204
+ "original_policy": {
1205
+ "mode": "parallel",
1206
+ "num_passes": 3,
1207
+ "feedback_loss_weight": 1.0,
1208
+ "attention_backend": "sdpa_auto",
1209
+ "checkpoint_refiner": false,
1210
+ "cache_layer_indices": null,
1211
+ "refiner_intermediate_size": 5632,
1212
+ "refiner_gate_bias": -2.0,
1213
+ "future_loss_weight": 0.0,
1214
+ "future_gradient_mode": "all",
1215
+ "draft_loss_weight": 0.0,
1216
+ "refined_future_loss_weight": 1.0,
1217
+ "pseudo_initializer": "A2",
1218
+ "pseudo_top_k": 128,
1219
+ "pseudo_temperature": 1.0,
1220
+ "refiner_variant": "R1",
1221
+ "allow_objective_change_on_resume": false
1222
+ },
1223
+ "status": "succeeded",
1224
+ "completed_tasks": 4,
1225
+ "node": "22.169.130.39",
1226
+ "gpu": 2,
1227
+ "result_dir": "/root/dl/runtime/script/loopTF-workspace/output/benchmarks/20260920-140758_r3_d0_p0_a2_r1_26703-prefix-full",
1228
+ "tasks": {
1229
+ "var_assign_d0_math_prefix_query": {
1230
+ "loss": 0.7666976825340996,
1231
+ "ppl": 2.1526457834604424,
1232
+ "accuracy": 0.788,
1233
+ "loss_sum": 1248.1838271655142,
1234
+ "num_tokens": 1628,
1235
+ "examples": 1000,
1236
+ "cropped_examples": 0
1237
+ },
1238
+ "var_assign_d0_code_prefix_query": {
1239
+ "loss": 0.2393993108986609,
1240
+ "ppl": 1.2704857541165646,
1241
+ "accuracy": 0.952,
1242
+ "loss_sum": 630.5777849070728,
1243
+ "num_tokens": 2634,
1244
+ "examples": 1000,
1245
+ "cropped_examples": 0
1246
+ },
1247
+ "var_assign_d1_math_prefix_query": {
1248
+ "loss": 2.0439432417521166,
1249
+ "ppl": 7.720995000078425,
1250
+ "accuracy": 0.159,
1251
+ "loss_sum": 3380.6821218580008,
1252
+ "num_tokens": 1654,
1253
+ "examples": 1000,
1254
+ "cropped_examples": 0
1255
+ },
1256
+ "var_assign_d1_code_prefix_query": {
1257
+ "loss": 1.2193537087728423,
1258
+ "ppl": 3.3849993311531907,
1259
+ "accuracy": 0.275,
1260
+ "loss_sum": 3241.042157918215,
1261
+ "num_tokens": 2658,
1262
+ "examples": 1000,
1263
+ "cropped_examples": 0
1264
+ }
1265
+ },
1266
+ "macro_loss": 1.0673484859894298,
1267
+ "macro_accuracy": 0.5435,
1268
+ "original_tasks": {}
1269
+ },
1270
+ {
1271
+ "name": "r3_d1_a0",
1272
+ "checkpoint_name": "r3_D1-A0-R0_7b_seed42_20260919-040211/step_26703",
1273
+ "checkpoint": "/taijifs_wzbd31/rentTest/hunyuan/kenanchen/loopTF-workspace/output/paper2604-six-final-20260919/checkpoints/r3_d1_a0/evaluation_native.pt",
1274
+ "checkpoint_sha256": "a4e7e9c38c48b4a71e612a53f9ac8f73c5acf7b0686b66a814a16ece1bd0ef36",
1275
+ "step": 26703,
1276
+ "architecture": "lookahead",
1277
+ "parameter_count": 661315072,
1278
+ "training_tokens": 7000031232,
1279
+ "inference": {
1280
+ "requested": "auto",
1281
+ "resolved": "fused",
1282
+ "full_bandwidth": true,
1283
+ "prompt_traversals": 2,
1284
+ "fused_prefill_passes": 1,
1285
+ "continuation": "one feedback transition per real token",
1286
+ "cache": "architecture-owned immutable real-prefix state"
1287
+ },
1288
+ "original_result_dir": "/root/dl/runtime/script/loopTF-workspace/output/benchmarks/20260920-000112_r3_d1_a0-full",
1289
+ "original_policy": {
1290
+ "mode": "parallel",
1291
+ "num_passes": 3,
1292
+ "feedback_loss_weight": 1.0,
1293
+ "attention_backend": "sdpa_auto",
1294
+ "checkpoint_refiner": false,
1295
+ "cache_layer_indices": null,
1296
+ "refiner_intermediate_size": 5632,
1297
+ "refiner_gate_bias": -2.0,
1298
+ "future_loss_weight": 0.2,
1299
+ "future_gradient_mode": "all",
1300
+ "draft_loss_weight": 0.2,
1301
+ "refined_future_loss_weight": 0.0,
1302
+ "pseudo_initializer": "A0",
1303
+ "pseudo_top_k": 128,
1304
+ "pseudo_temperature": 1.0,
1305
+ "refiner_variant": "R0",
1306
+ "allow_objective_change_on_resume": false
1307
+ },
1308
+ "status": "succeeded",
1309
+ "completed_tasks": 4,
1310
+ "node": "22.169.130.39",
1311
+ "gpu": 3,
1312
+ "result_dir": "/root/dl/runtime/script/loopTF-workspace/output/benchmarks/20260920-140827_r3_d1_a0-prefix-full",
1313
+ "tasks": {
1314
+ "var_assign_d0_math_prefix_query": {
1315
+ "loss": 0.801322063669123,
1316
+ "ppl": 2.2284851811140425,
1317
+ "accuracy": 0.704,
1318
+ "loss_sum": 1304.5523196533322,
1319
+ "num_tokens": 1628,
1320
+ "examples": 1000,
1321
+ "cropped_examples": 0
1322
+ },
1323
+ "var_assign_d0_code_prefix_query": {
1324
+ "loss": 0.27040091631469093,
1325
+ "ppl": 1.310489742145124,
1326
+ "accuracy": 0.92,
1327
+ "loss_sum": 712.2360135728959,
1328
+ "num_tokens": 2634,
1329
+ "examples": 1000,
1330
+ "cropped_examples": 0
1331
+ },
1332
+ "var_assign_d1_math_prefix_query": {
1333
+ "loss": 2.3334992481654617,
1334
+ "ppl": 10.313969599908269,
1335
+ "accuracy": 0.09,
1336
+ "loss_sum": 3859.6077564656734,
1337
+ "num_tokens": 1654,
1338
+ "examples": 1000,
1339
+ "cropped_examples": 0
1340
+ },
1341
+ "var_assign_d1_code_prefix_query": {
1342
+ "loss": 1.2137934214598798,
1343
+ "ppl": 3.366229992133896,
1344
+ "accuracy": 0.237,
1345
+ "loss_sum": 3226.2629142403603,
1346
+ "num_tokens": 2658,
1347
+ "examples": 1000,
1348
+ "cropped_examples": 0
1349
+ }
1350
+ },
1351
+ "macro_loss": 1.1547539124022888,
1352
+ "macro_accuracy": 0.48775,
1353
+ "original_tasks": {
1354
+ "var_assign_d0_math": {
1355
+ "loss": 0.5126674081332828,
1356
+ "ppl": 1.6697391358231228,
1357
+ "accuracy": 0.506,
1358
+ "loss_sum": 1347.2899485742673,
1359
+ "num_tokens": 2628,
1360
+ "examples": 1000,
1361
+ "cropped_examples": 0
1362
+ },
1363
+ "var_assign_d0_code": {
1364
+ "loss": 0.33868700629008275,
1365
+ "ppl": 1.4031041137160118,
1366
+ "accuracy": 0.707,
1367
+ "loss_sum": 892.1015745680779,
1368
+ "num_tokens": 2634,
1369
+ "examples": 1000,
1370
+ "cropped_examples": 0
1371
+ },
1372
+ "var_assign_d1_math": {
1373
+ "loss": 0.9369979491464361,
1374
+ "ppl": 2.5523077480645755,
1375
+ "accuracy": 0.2,
1376
+ "loss_sum": 2486.7925570346415,
1377
+ "num_tokens": 2654,
1378
+ "examples": 1000,
1379
+ "cropped_examples": 0
1380
+ },
1381
+ "var_assign_d1_code": {
1382
+ "loss": 0.8713010571430317,
1383
+ "ppl": 2.3900183820377636,
1384
+ "accuracy": 0.254,
1385
+ "loss_sum": 2315.9182098861784,
1386
+ "num_tokens": 2658,
1387
+ "examples": 1000,
1388
+ "cropped_examples": 0
1389
+ }
1390
+ }
1391
+ },
1392
+ {
1393
+ "name": "vanilla",
1394
+ "checkpoint_name": "20260916-204226_vanilla_476m_7b",
1395
+ "checkpoint": "/taijifs_wzbd31/rentTest/hunyuan/kenanchen/loopTF-workspace/output/evaluation-readiness/checkpoints/vanilla/evaluation_native.pt",
1396
+ "checkpoint_sha256": "dd4edf8f146d3215190bd526faf3e9085165c3dfecb95cab1e1471b9e6479390",
1397
+ "step": 26703,
1398
+ "architecture": "paper2604_vanilla",
1399
+ "parameter_count": 566144128,
1400
+ "training_tokens": 7000031232,
1401
+ "inference": {
1402
+ "requested": "auto",
1403
+ "resolved": "fixed-pass",
1404
+ "full_bandwidth": false,
1405
+ "prompt_traversals": null,
1406
+ "fused_prefill_passes": 0,
1407
+ "continuation": "fixed-pass",
1408
+ "cache": "none"
1409
+ },
1410
+ "original_result_dir": "/root/dl/runtime/script/loopTF-workspace/output/benchmarks/20260920-000106_vanilla-full",
1411
+ "original_policy": {
1412
+ "num_passes": 1
1413
+ },
1414
+ "status": "succeeded",
1415
+ "completed_tasks": 4,
1416
+ "node": "22.169.167.168",
1417
+ "gpu": 0,
1418
+ "result_dir": "/root/dl/runtime/script/loopTF-workspace/output/benchmarks/20260920-140827_vanilla-prefix-full",
1419
+ "tasks": {
1420
+ "var_assign_d0_math_prefix_query": {
1421
+ "loss": 1.024175554372059,
1422
+ "ppl": 2.784798598847634,
1423
+ "accuracy": 0.704,
1424
+ "loss_sum": 1667.3578025177121,
1425
+ "num_tokens": 1628,
1426
+ "examples": 1000,
1427
+ "cropped_examples": 0
1428
+ },
1429
+ "var_assign_d0_code_prefix_query": {
1430
+ "loss": 0.5120287581814358,
1431
+ "ppl": 1.6686730974533432,
1432
+ "accuracy": 0.795,
1433
+ "loss_sum": 1348.683749049902,
1434
+ "num_tokens": 2634,
1435
+ "examples": 1000,
1436
+ "cropped_examples": 0
1437
+ },
1438
+ "var_assign_d1_math_prefix_query": {
1439
+ "loss": 2.5404749165999703,
1440
+ "ppl": 12.685694187209599,
1441
+ "accuracy": 0.032,
1442
+ "loss_sum": 4201.945512056351,
1443
+ "num_tokens": 1654,
1444
+ "examples": 1000,
1445
+ "cropped_examples": 0
1446
+ },
1447
+ "var_assign_d1_code_prefix_query": {
1448
+ "loss": 1.1742466776284892,
1449
+ "ppl": 3.23570449689985,
1450
+ "accuracy": 0.25,
1451
+ "loss_sum": 3121.147669136524,
1452
+ "num_tokens": 2658,
1453
+ "examples": 1000,
1454
+ "cropped_examples": 0
1455
+ }
1456
+ },
1457
+ "macro_loss": 1.3127314766954885,
1458
+ "macro_accuracy": 0.44525,
1459
+ "original_tasks": {
1460
+ "var_assign_d0_math": {
1461
+ "loss": 0.6255110967025456,
1462
+ "ppl": 1.869201055833641,
1463
+ "accuracy": 0.398,
1464
+ "loss_sum": 1643.8431621342897,
1465
+ "num_tokens": 2628,
1466
+ "examples": 1000,
1467
+ "cropped_examples": 0
1468
+ },
1469
+ "var_assign_d0_code": {
1470
+ "loss": 0.654858390705189,
1471
+ "ppl": 1.9248699186037819,
1472
+ "accuracy": 0.454,
1473
+ "loss_sum": 1724.8970011174679,
1474
+ "num_tokens": 2634,
1475
+ "examples": 1000,
1476
+ "cropped_examples": 0
1477
+ },
1478
+ "var_assign_d1_math": {
1479
+ "loss": 0.8273450185817799,
1480
+ "ppl": 2.287238097191743,
1481
+ "accuracy": 0.26,
1482
+ "loss_sum": 2195.773679316044,
1483
+ "num_tokens": 2654,
1484
+ "examples": 1000,
1485
+ "cropped_examples": 0
1486
+ },
1487
+ "var_assign_d1_code": {
1488
+ "loss": 0.8524075931875754,
1489
+ "ppl": 2.345286556080457,
1490
+ "accuracy": 0.273,
1491
+ "loss_sum": 2265.6993826925755,
1492
+ "num_tokens": 2658,
1493
+ "examples": 1000,
1494
+ "cropped_examples": 0
1495
+ }
1496
+ }
1497
+ },
1498
+ {
1499
+ "name": "alpha0_seed43_24000",
1500
+ "checkpoint_name": "r3_b300_alpha0p0-seed43_product_20260918-170109/step_24000",
1501
+ "checkpoint": "/taijifs_wzbd31/rentTest/hunyuan/kenanchen/loopTF-workspace/output/paper2604-expanded-20260920/checkpoints/alpha0_seed43_24000/evaluation_native.pt",
1502
+ "checkpoint_sha256": "bc57d48a8e45c73167afd7c038b071d7757839a0372197f88076c1f2361f54d2",
1503
+ "step": 24000,
1504
+ "architecture": "lookahead",
1505
+ "parameter_count": 661315072,
1506
+ "training_tokens": 6291456000,
1507
+ "inference": {
1508
+ "requested": "auto",
1509
+ "resolved": "fused",
1510
+ "full_bandwidth": true,
1511
+ "prompt_traversals": 2,
1512
+ "fused_prefill_passes": 1,
1513
+ "continuation": "one feedback transition per real token",
1514
+ "cache": "architecture-owned immutable real-prefix state"
1515
+ },
1516
+ "original_result_dir": "/root/dl/runtime/script/loopTF-workspace/output/benchmarks/20260920-103100_alpha0_seed43_24000-full",
1517
+ "original_policy": {
1518
+ "mode": "parallel",
1519
+ "num_passes": 3,
1520
+ "feedback_loss_weight": 1.0,
1521
+ "attention_backend": "sdpa_auto",
1522
+ "checkpoint_refiner": false,
1523
+ "cache_layer_indices": null,
1524
+ "refiner_intermediate_size": 5632,
1525
+ "refiner_gate_bias": -2.0,
1526
+ "future_loss_weight": 0.0,
1527
+ "future_gradient_mode": "all",
1528
+ "draft_loss_weight": 0.0,
1529
+ "refined_future_loss_weight": 0.0,
1530
+ "pseudo_initializer": "A0",
1531
+ "pseudo_top_k": 128,
1532
+ "pseudo_temperature": 1.0,
1533
+ "refiner_variant": "R0",
1534
+ "allow_objective_change_on_resume": false
1535
+ },
1536
+ "status": "succeeded",
1537
+ "completed_tasks": 4,
1538
+ "node": "22.169.167.168",
1539
+ "gpu": 1,
1540
+ "result_dir": "/root/dl/runtime/script/loopTF-workspace/output/benchmarks/20260920-140827_alpha0_seed43_24000-prefix-full",
1541
+ "tasks": {
1542
+ "var_assign_d0_math_prefix_query": {
1543
+ "loss": 0.4682283340324018,
1544
+ "ppl": 1.5971620475057635,
1545
+ "accuracy": 0.949,
1546
+ "loss_sum": 762.2757278047502,
1547
+ "num_tokens": 1628,
1548
+ "examples": 1000,
1549
+ "cropped_examples": 0
1550
+ },
1551
+ "var_assign_d0_code_prefix_query": {
1552
+ "loss": 0.25010449535681195,
1553
+ "ppl": 1.2841595983923744,
1554
+ "accuracy": 0.943,
1555
+ "loss_sum": 658.7752407698426,
1556
+ "num_tokens": 2634,
1557
+ "examples": 1000,
1558
+ "cropped_examples": 0
1559
+ },
1560
+ "var_assign_d1_math_prefix_query": {
1561
+ "loss": 2.396551696887688,
1562
+ "ppl": 10.98523058911664,
1563
+ "accuracy": 0.077,
1564
+ "loss_sum": 3963.896506652236,
1565
+ "num_tokens": 1654,
1566
+ "examples": 1000,
1567
+ "cropped_examples": 0
1568
+ },
1569
+ "var_assign_d1_code_prefix_query": {
1570
+ "loss": 1.119192756471372,
1571
+ "ppl": 3.0623811178965936,
1572
+ "accuracy": 0.241,
1573
+ "loss_sum": 2974.8143467009068,
1574
+ "num_tokens": 2658,
1575
+ "examples": 1000,
1576
+ "cropped_examples": 0
1577
+ }
1578
+ },
1579
+ "macro_loss": 1.0585193206870684,
1580
+ "macro_accuracy": 0.5525,
1581
+ "original_tasks": {
1582
+ "var_assign_d0_math": {
1583
+ "loss": 0.24017553337623934,
1584
+ "ppl": 1.2714723165627908,
1585
+ "accuracy": 0.803,
1586
+ "loss_sum": 631.181301712757,
1587
+ "num_tokens": 2628,
1588
+ "examples": 1000,
1589
+ "cropped_examples": 0
1590
+ },
1591
+ "var_assign_d0_code": {
1592
+ "loss": 0.22229844175725355,
1593
+ "ppl": 1.2489440593094363,
1594
+ "accuracy": 0.834,
1595
+ "loss_sum": 585.5340955886059,
1596
+ "num_tokens": 2634,
1597
+ "examples": 1000,
1598
+ "cropped_examples": 0
1599
+ },
1600
+ "var_assign_d1_math": {
1601
+ "loss": 0.7958082678075459,
1602
+ "ppl": 2.216231581712088,
1603
+ "accuracy": 0.258,
1604
+ "loss_sum": 2112.0751427612267,
1605
+ "num_tokens": 2654,
1606
+ "examples": 1000,
1607
+ "cropped_examples": 0
1608
+ },
1609
+ "var_assign_d1_code": {
1610
+ "loss": 0.7908044640674551,
1611
+ "ppl": 2.2051696926761446,
1612
+ "accuracy": 0.269,
1613
+ "loss_sum": 2101.9582654912956,
1614
+ "num_tokens": 2658,
1615
+ "examples": 1000,
1616
+ "cropped_examples": 0
1617
+ }
1618
+ }
1619
+ },
1620
+ {
1621
+ "name": "fbw_l2_l11_l20_24000",
1622
+ "checkpoint_name": "fbw_L2L11L20_7b_seed42_20260920-023045/step_24000",
1623
+ "checkpoint": "/taijifs_wzbd31/rentTest/hunyuan/kenanchen/loopTF-workspace/output/paper2604-all-20260920/checkpoints/fbw_l2_l11_l20_24000/evaluation_native.pt",
1624
+ "checkpoint_sha256": "aceef5da3dec7eecc7d9c71f0240225ab78385f01e4828f742b38e46fdac5eca",
1625
+ "step": 24000,
1626
+ "architecture": "full_bandwidth",
1627
+ "parameter_count": 576059264,
1628
+ "training_tokens": 6291456000,
1629
+ "inference": {
1630
+ "requested": "auto",
1631
+ "resolved": "fused",
1632
+ "full_bandwidth": true,
1633
+ "prompt_traversals": 2,
1634
+ "fused_prefill_passes": 1,
1635
+ "continuation": "one feedback transition per real token",
1636
+ "cache": "architecture-owned immutable real-prefix state"
1637
+ },
1638
+ "original_result_dir": "/root/dl/runtime/script/loopTF-workspace/output/benchmarks/20260920-103146_fbw_l2_l11_l20_24000-full",
1639
+ "original_policy": {
1640
+ "num_passes": 3,
1641
+ "depth_loops": 1,
1642
+ "feedback_layers": [
1643
+ 2,
1644
+ 11,
1645
+ 20
1646
+ ],
1647
+ "depth_share_kv": false,
1648
+ "final_adapter": {
1649
+ "kind": "none",
1650
+ "ffn_size": null,
1651
+ "attention_backend": "sdpa_auto",
1652
+ "checkpoint_refiner": true
1653
+ }
1654
+ },
1655
+ "status": "succeeded",
1656
+ "completed_tasks": 4,
1657
+ "node": "22.169.167.168",
1658
+ "gpu": 5,
1659
+ "result_dir": "/root/dl/runtime/script/loopTF-workspace/output/benchmarks/20260920-140829_fbw_l2_l11_l20_24000-prefix-full",
1660
+ "tasks": {
1661
+ "var_assign_d0_math_prefix_query": {
1662
+ "loss": 0.771849320889096,
1663
+ "ppl": 2.1637640500983526,
1664
+ "accuracy": 0.811,
1665
+ "loss_sum": 1256.5706944074482,
1666
+ "num_tokens": 1628,
1667
+ "examples": 1000,
1668
+ "cropped_examples": 0
1669
+ },
1670
+ "var_assign_d0_code_prefix_query": {
1671
+ "loss": 0.24532383162918625,
1672
+ "ppl": 1.278035114393565,
1673
+ "accuracy": 0.903,
1674
+ "loss_sum": 646.1829725112766,
1675
+ "num_tokens": 2634,
1676
+ "examples": 1000,
1677
+ "cropped_examples": 0
1678
+ },
1679
+ "var_assign_d1_math_prefix_query": {
1680
+ "loss": 2.5469410528655207,
1681
+ "ppl": 12.767987387240439,
1682
+ "accuracy": 0.054,
1683
+ "loss_sum": 4212.640501439571,
1684
+ "num_tokens": 1654,
1685
+ "examples": 1000,
1686
+ "cropped_examples": 0
1687
+ },
1688
+ "var_assign_d1_code_prefix_query": {
1689
+ "loss": 1.0375313504246548,
1690
+ "ppl": 2.822241282831491,
1691
+ "accuracy": 0.28,
1692
+ "loss_sum": 2757.7583294287324,
1693
+ "num_tokens": 2658,
1694
+ "examples": 1000,
1695
+ "cropped_examples": 0
1696
+ }
1697
+ },
1698
+ "macro_loss": 1.1504113889521144,
1699
+ "macro_accuracy": 0.512,
1700
+ "original_tasks": {
1701
+ "var_assign_d0_math": {
1702
+ "loss": 0.7188502418538598,
1703
+ "ppl": 2.052072466726588,
1704
+ "accuracy": 0.325,
1705
+ "loss_sum": 1889.1384355919436,
1706
+ "num_tokens": 2628,
1707
+ "examples": 1000,
1708
+ "cropped_examples": 0
1709
+ },
1710
+ "var_assign_d0_code": {
1711
+ "loss": 0.625273352526001,
1712
+ "ppl": 1.8687567169894048,
1713
+ "accuracy": 0.431,
1714
+ "loss_sum": 1646.9700105534866,
1715
+ "num_tokens": 2634,
1716
+ "examples": 1000,
1717
+ "cropped_examples": 0
1718
+ },
1719
+ "var_assign_d1_math": {
1720
+ "loss": 0.966856686145028,
1721
+ "ppl": 2.629665590344807,
1722
+ "accuracy": 0.198,
1723
+ "loss_sum": 2566.037645028904,
1724
+ "num_tokens": 2654,
1725
+ "examples": 1000,
1726
+ "cropped_examples": 0
1727
+ },
1728
+ "var_assign_d1_code": {
1729
+ "loss": 0.9253820733547748,
1730
+ "ppl": 2.522831983119354,
1731
+ "accuracy": 0.23,
1732
+ "loss_sum": 2459.6655509769917,
1733
+ "num_tokens": 2658,
1734
+ "examples": 1000,
1735
+ "cropped_examples": 0
1736
+ }
1737
+ }
1738
+ },
1739
+ {
1740
+ "name": "loop_adapter_24000",
1741
+ "checkpoint_name": "paper2604_loop2_476m_7b_online_FBW0_K3_adapter_norm/step_24000",
1742
+ "checkpoint": "/taijifs_wzbd31/rentTest/hunyuan/kenanchen/loopTF-workspace/output/paper2604-expanded-20260920/checkpoints/loop_adapter_24000/evaluation_native.pt",
1743
+ "checkpoint_sha256": "bce3588669a67446d11ecf8434bf867d337d82f4f44dd4cd73fb0f4875026297",
1744
+ "step": 24000,
1745
+ "architecture": "full_bandwidth_loop",
1746
+ "parameter_count": 653383808,
1747
+ "training_tokens": 6291456000,
1748
+ "inference": {
1749
+ "requested": "auto",
1750
+ "resolved": "fused",
1751
+ "full_bandwidth": true,
1752
+ "prompt_traversals": 2,
1753
+ "fused_prefill_passes": 1,
1754
+ "continuation": "one feedback transition per real token",
1755
+ "cache": "architecture-owned immutable real-prefix state"
1756
+ },
1757
+ "original_result_dir": "/root/dl/runtime/script/loopTF-workspace/output/benchmarks/20260920-084127_loop_adapter_24000-full",
1758
+ "original_policy": {
1759
+ "num_passes": 3,
1760
+ "depth_loops": 2,
1761
+ "feedback_layers": [
1762
+ 20
1763
+ ],
1764
+ "depth_share_kv": false,
1765
+ "final_adapter": {
1766
+ "kind": "lookahead_refiner",
1767
+ "ffn_size": 5632,
1768
+ "attention_backend": "sdpa_auto",
1769
+ "checkpoint_refiner": true
1770
+ }
1771
+ },
1772
+ "status": "succeeded",
1773
+ "completed_tasks": 4,
1774
+ "node": "22.169.167.168",
1775
+ "gpu": 6,
1776
+ "result_dir": "/root/dl/runtime/script/loopTF-workspace/output/benchmarks/20260920-140831_loop_adapter_24000-prefix-full",
1777
+ "tasks": {
1778
+ "var_assign_d0_math_prefix_query": {
1779
+ "loss": 0.7688110933069332,
1780
+ "ppl": 2.157200019039444,
1781
+ "accuracy": 0.763,
1782
+ "loss_sum": 1251.6244599036872,
1783
+ "num_tokens": 1628,
1784
+ "examples": 1000,
1785
+ "cropped_examples": 0
1786
+ },
1787
+ "var_assign_d0_code_prefix_query": {
1788
+ "loss": 0.3715563605445096,
1789
+ "ppl": 1.4499895659999795,
1790
+ "accuracy": 0.853,
1791
+ "loss_sum": 978.6794536742382,
1792
+ "num_tokens": 2634,
1793
+ "examples": 1000,
1794
+ "cropped_examples": 0
1795
+ },
1796
+ "var_assign_d1_math_prefix_query": {
1797
+ "loss": 2.6806160993648844,
1798
+ "ppl": 14.594081931259112,
1799
+ "accuracy": 0.029,
1800
+ "loss_sum": 4433.739028349519,
1801
+ "num_tokens": 1654,
1802
+ "examples": 1000,
1803
+ "cropped_examples": 0
1804
+ },
1805
+ "var_assign_d1_code_prefix_query": {
1806
+ "loss": 1.1632124997641782,
1807
+ "ppl": 3.2001974139563583,
1808
+ "accuracy": 0.213,
1809
+ "loss_sum": 3091.8188243731856,
1810
+ "num_tokens": 2658,
1811
+ "examples": 1000,
1812
+ "cropped_examples": 0
1813
+ }
1814
+ },
1815
+ "macro_loss": 1.2460490132451263,
1816
+ "macro_accuracy": 0.4645,
1817
+ "original_tasks": {
1818
+ "var_assign_d0_math": {
1819
+ "loss": 0.24550611267126446,
1820
+ "ppl": 1.2782680971995586,
1821
+ "accuracy": 0.816,
1822
+ "loss_sum": 645.190064100083,
1823
+ "num_tokens": 2628,
1824
+ "examples": 1000,
1825
+ "cropped_examples": 0
1826
+ },
1827
+ "var_assign_d0_code": {
1828
+ "loss": 0.23753670796938664,
1829
+ "ppl": 1.2681215461032416,
1830
+ "accuracy": 0.932,
1831
+ "loss_sum": 625.6716887913644,
1832
+ "num_tokens": 2634,
1833
+ "examples": 1000,
1834
+ "cropped_examples": 0
1835
+ },
1836
+ "var_assign_d1_math": {
1837
+ "loss": 0.7843343140817084,
1838
+ "ppl": 2.1909479719669633,
1839
+ "accuracy": 0.285,
1840
+ "loss_sum": 2081.623269572854,
1841
+ "num_tokens": 2654,
1842
+ "examples": 1000,
1843
+ "cropped_examples": 0
1844
+ },
1845
+ "var_assign_d1_code": {
1846
+ "loss": 0.8455226423430479,
1847
+ "ppl": 2.3291948323546157,
1848
+ "accuracy": 0.283,
1849
+ "loss_sum": 2247.3991833478212,
1850
+ "num_tokens": 2658,
1851
+ "examples": 1000,
1852
+ "cropped_examples": 0
1853
+ }
1854
+ }
1855
+ },
1856
+ {
1857
+ "name": "loop_sharekv_24000",
1858
+ "checkpoint_name": "paper2604_loop2_476m_7b_online_FBW0_K3_adapter_norm_sharekv_20260918-043531/step_24000",
1859
+ "checkpoint": "/taijifs_wzbd31/rentTest/hunyuan/kenanchen/loopTF-workspace/output/paper2604-expanded-20260920/checkpoints/loop_sharekv_24000/evaluation_native.pt",
1860
+ "checkpoint_sha256": "402201b77fd972dc99901eeb02f0ab9473031f1ab89272bca50bfa2dd1e27dac",
1861
+ "step": 24000,
1862
+ "architecture": "full_bandwidth_loop",
1863
+ "parameter_count": 653383808,
1864
+ "training_tokens": 6291456000,
1865
+ "inference": {
1866
+ "requested": "auto",
1867
+ "resolved": "fused",
1868
+ "full_bandwidth": true,
1869
+ "prompt_traversals": 2,
1870
+ "fused_prefill_passes": 1,
1871
+ "continuation": "one feedback transition per real token",
1872
+ "cache": "architecture-owned immutable real-prefix state"
1873
+ },
1874
+ "original_result_dir": "/root/dl/runtime/script/loopTF-workspace/output/benchmarks/20260920-084129_loop_sharekv_24000-full",
1875
+ "original_policy": {
1876
+ "num_passes": 3,
1877
+ "depth_loops": 2,
1878
+ "feedback_layers": [
1879
+ 20
1880
+ ],
1881
+ "depth_share_kv": true,
1882
+ "final_adapter": {
1883
+ "kind": "lookahead_refiner",
1884
+ "ffn_size": 5632,
1885
+ "attention_backend": "sdpa_auto",
1886
+ "checkpoint_refiner": true
1887
+ }
1888
+ },
1889
+ "status": "succeeded",
1890
+ "completed_tasks": 4,
1891
+ "node": "22.169.167.168",
1892
+ "gpu": 7,
1893
+ "result_dir": "/root/dl/runtime/script/loopTF-workspace/output/benchmarks/20260920-140905_loop_sharekv_24000-prefix-full",
1894
+ "tasks": {
1895
+ "var_assign_d0_math_prefix_query": {
1896
+ "loss": 1.0303520365350443,
1897
+ "ppl": 2.802052085797982,
1898
+ "accuracy": 0.669,
1899
+ "loss_sum": 1677.413115479052,
1900
+ "num_tokens": 1628,
1901
+ "examples": 1000,
1902
+ "cropped_examples": 0
1903
+ },
1904
+ "var_assign_d0_code_prefix_query": {
1905
+ "loss": 0.2604632708846433,
1906
+ "ppl": 1.2975310558093471,
1907
+ "accuracy": 0.907,
1908
+ "loss_sum": 686.0602555101505,
1909
+ "num_tokens": 2634,
1910
+ "examples": 1000,
1911
+ "cropped_examples": 0
1912
+ },
1913
+ "var_assign_d1_math_prefix_query": {
1914
+ "loss": 2.261107094730042,
1915
+ "ppl": 9.593704429244776,
1916
+ "accuracy": 0.078,
1917
+ "loss_sum": 3739.87113468349,
1918
+ "num_tokens": 1654,
1919
+ "examples": 1000,
1920
+ "cropped_examples": 0
1921
+ },
1922
+ "var_assign_d1_code_prefix_query": {
1923
+ "loss": 1.0652912875330476,
1924
+ "ppl": 2.901684085368831,
1925
+ "accuracy": 0.285,
1926
+ "loss_sum": 2831.5442422628403,
1927
+ "num_tokens": 2658,
1928
+ "examples": 1000,
1929
+ "cropped_examples": 0
1930
+ }
1931
+ },
1932
+ "macro_loss": 1.1543034224206943,
1933
+ "macro_accuracy": 0.48475,
1934
+ "original_tasks": {
1935
+ "var_assign_d0_math": {
1936
+ "loss": 0.45675739788420533,
1937
+ "ppl": 1.57894578219315,
1938
+ "accuracy": 0.633,
1939
+ "loss_sum": 1200.3584416396916,
1940
+ "num_tokens": 2628,
1941
+ "examples": 1000,
1942
+ "cropped_examples": 0
1943
+ },
1944
+ "var_assign_d0_code": {
1945
+ "loss": 0.42422725412437384,
1946
+ "ppl": 1.5284088915444713,
1947
+ "accuracy": 0.671,
1948
+ "loss_sum": 1117.4145873636007,
1949
+ "num_tokens": 2634,
1950
+ "examples": 1000,
1951
+ "cropped_examples": 0
1952
+ },
1953
+ "var_assign_d1_math": {
1954
+ "loss": 0.8732526009781644,
1955
+ "ppl": 2.3946871618595593,
1956
+ "accuracy": 0.209,
1957
+ "loss_sum": 2317.6124029960483,
1958
+ "num_tokens": 2654,
1959
+ "examples": 1000,
1960
+ "cropped_examples": 0
1961
+ },
1962
+ "var_assign_d1_code": {
1963
+ "loss": 0.8642199649731296,
1964
+ "ppl": 2.373154220394927,
1965
+ "accuracy": 0.267,
1966
+ "loss_sum": 2297.0966668985784,
1967
+ "num_tokens": 2658,
1968
+ "examples": 1000,
1969
+ "cropped_examples": 0
1970
+ }
1971
+ }
1972
+ },
1973
+ {
1974
+ "name": "r3_d0_a2_24000",
1975
+ "checkpoint_name": "r3_D0-A2-R0_7b_seed42_20260919-040210/step_24000",
1976
+ "checkpoint": "/taijifs_wzbd31/rentTest/hunyuan/kenanchen/loopTF-workspace/output/paper2604-expanded-20260920/checkpoints/r3_d0_a2_24000/evaluation_native.pt",
1977
+ "checkpoint_sha256": "3df6a65722f83940016fa47ae3c7c756b784253127a1fe21f28a542b54d2ffe4",
1978
+ "step": 24000,
1979
+ "architecture": "lookahead",
1980
+ "parameter_count": 659332608,
1981
+ "training_tokens": 6291456000,
1982
+ "inference": {
1983
+ "requested": "auto",
1984
+ "resolved": "fused",
1985
+ "full_bandwidth": true,
1986
+ "prompt_traversals": 2,
1987
+ "fused_prefill_passes": 1,
1988
+ "continuation": "one feedback transition per real token",
1989
+ "cache": "architecture-owned immutable real-prefix state"
1990
+ },
1991
+ "original_result_dir": "/root/dl/runtime/script/loopTF-workspace/output/benchmarks/20260920-084130_r3_d0_a2_24000-full",
1992
+ "original_policy": {
1993
+ "mode": "parallel",
1994
+ "num_passes": 3,
1995
+ "feedback_loss_weight": 1.0,
1996
+ "attention_backend": "sdpa_auto",
1997
+ "checkpoint_refiner": false,
1998
+ "cache_layer_indices": null,
1999
+ "refiner_intermediate_size": 5632,
2000
+ "refiner_gate_bias": -2.0,
2001
+ "future_loss_weight": 0.2,
2002
+ "future_gradient_mode": "all",
2003
+ "draft_loss_weight": 0.0,
2004
+ "refined_future_loss_weight": 0.0,
2005
+ "pseudo_initializer": "A2",
2006
+ "pseudo_top_k": 128,
2007
+ "pseudo_temperature": 1.0,
2008
+ "refiner_variant": "R0",
2009
+ "allow_objective_change_on_resume": false
2010
+ },
2011
+ "status": "succeeded",
2012
+ "completed_tasks": 4,
2013
+ "node": "22.169.167.168",
2014
+ "gpu": 0,
2015
+ "result_dir": "/root/dl/runtime/script/loopTF-workspace/output/benchmarks/20260920-141038_r3_d0_a2_24000-prefix-full",
2016
+ "tasks": {
2017
+ "var_assign_d0_math_prefix_query": {
2018
+ "loss": 0.9568242601231106,
2019
+ "ppl": 2.6034155607580596,
2020
+ "accuracy": 0.643,
2021
+ "loss_sum": 1557.7098954804242,
2022
+ "num_tokens": 1628,
2023
+ "examples": 1000,
2024
+ "cropped_examples": 0
2025
+ },
2026
+ "var_assign_d0_code_prefix_query": {
2027
+ "loss": 0.27780481549892105,
2028
+ "ppl": 1.3202284839211493,
2029
+ "accuracy": 0.86,
2030
+ "loss_sum": 731.7378840241581,
2031
+ "num_tokens": 2634,
2032
+ "examples": 1000,
2033
+ "cropped_examples": 0
2034
+ },
2035
+ "var_assign_d1_math_prefix_query": {
2036
+ "loss": 2.8758340317580258,
2037
+ "ppl": 17.74021385479974,
2038
+ "accuracy": 0.02,
2039
+ "loss_sum": 4756.629488527775,
2040
+ "num_tokens": 1654,
2041
+ "examples": 1000,
2042
+ "cropped_examples": 0
2043
+ },
2044
+ "var_assign_d1_code_prefix_query": {
2045
+ "loss": 1.241245138305439,
2046
+ "ppl": 3.4599188613072753,
2047
+ "accuracy": 0.239,
2048
+ "loss_sum": 3299.229577615857,
2049
+ "num_tokens": 2658,
2050
+ "examples": 1000,
2051
+ "cropped_examples": 0
2052
+ }
2053
+ },
2054
+ "macro_loss": 1.337927061421374,
2055
+ "macro_accuracy": 0.4405,
2056
+ "original_tasks": {
2057
+ "var_assign_d0_math": {
2058
+ "loss": 0.5420153469986266,
2059
+ "ppl": 1.719468698693371,
2060
+ "accuracy": 0.534,
2061
+ "loss_sum": 1424.416331912391,
2062
+ "num_tokens": 2628,
2063
+ "examples": 1000,
2064
+ "cropped_examples": 0
2065
+ },
2066
+ "var_assign_d0_code": {
2067
+ "loss": 0.4673232916833079,
2068
+ "ppl": 1.5957172021361519,
2069
+ "accuracy": 0.661,
2070
+ "loss_sum": 1230.929550293833,
2071
+ "num_tokens": 2634,
2072
+ "examples": 1000,
2073
+ "cropped_examples": 0
2074
+ },
2075
+ "var_assign_d1_math": {
2076
+ "loss": 0.886408140112947,
2077
+ "ppl": 2.4263986963458875,
2078
+ "accuracy": 0.234,
2079
+ "loss_sum": 2352.5272038597614,
2080
+ "num_tokens": 2654,
2081
+ "examples": 1000,
2082
+ "cropped_examples": 0
2083
+ },
2084
+ "var_assign_d1_code": {
2085
+ "loss": 0.9103312402307718,
2086
+ "ppl": 2.485145577259183,
2087
+ "accuracy": 0.255,
2088
+ "loss_sum": 2419.6604365333915,
2089
+ "num_tokens": 2658,
2090
+ "examples": 1000,
2091
+ "cropped_examples": 0
2092
+ }
2093
+ }
2094
+ },
2095
+ {
2096
+ "name": "r3_d0_p0_a0_r1_24000",
2097
+ "checkpoint_name": "r3_D0-P0-A0-R1_7b_seed42_20260919-161651/step_24000",
2098
+ "checkpoint": "/taijifs_wzbd31/rentTest/hunyuan/kenanchen/loopTF-workspace/output/paper2604-all-20260920/checkpoints/r3_d0_p0_a0_r1_24000/evaluation_native.pt",
2099
+ "checkpoint_sha256": "20decd269c0bda7af6b779a355785d5b61e6d18f804d52c6c0d49b7fccc8aa93",
2100
+ "step": 24000,
2101
+ "architecture": "lookahead",
2102
+ "parameter_count": 667258240,
2103
+ "training_tokens": 6291456000,
2104
+ "inference": {
2105
+ "requested": "auto",
2106
+ "resolved": "fused",
2107
+ "full_bandwidth": true,
2108
+ "prompt_traversals": 2,
2109
+ "fused_prefill_passes": 1,
2110
+ "continuation": "one feedback transition per real token",
2111
+ "cache": "architecture-owned immutable real-prefix state"
2112
+ },
2113
+ "original_result_dir": "/root/dl/runtime/script/loopTF-workspace/output/benchmarks/20260920-103331_r3_d0_p0_a0_r1_24000-full",
2114
+ "original_policy": {
2115
+ "mode": "parallel",
2116
+ "num_passes": 3,
2117
+ "feedback_loss_weight": 1.0,
2118
+ "attention_backend": "sdpa_auto",
2119
+ "checkpoint_refiner": false,
2120
+ "cache_layer_indices": null,
2121
+ "refiner_intermediate_size": 5632,
2122
+ "refiner_gate_bias": -2.0,
2123
+ "future_loss_weight": 0.0,
2124
+ "future_gradient_mode": "all",
2125
+ "draft_loss_weight": 0.0,
2126
+ "refined_future_loss_weight": 1.0,
2127
+ "pseudo_initializer": "A0",
2128
+ "pseudo_top_k": 128,
2129
+ "pseudo_temperature": 1.0,
2130
+ "refiner_variant": "R1",
2131
+ "allow_objective_change_on_resume": false
2132
+ },
2133
+ "status": "succeeded",
2134
+ "completed_tasks": 4,
2135
+ "node": "22.169.167.168",
2136
+ "gpu": 4,
2137
+ "result_dir": "/root/dl/runtime/script/loopTF-workspace/output/benchmarks/20260920-141240_r3_d0_p0_a0_r1_24000-prefix-full",
2138
+ "tasks": {
2139
+ "var_assign_d0_math_prefix_query": {
2140
+ "loss": 0.7848784902455692,
2141
+ "ppl": 2.1921405580886075,
2142
+ "accuracy": 0.736,
2143
+ "loss_sum": 1277.7821821197867,
2144
+ "num_tokens": 1628,
2145
+ "examples": 1000,
2146
+ "cropped_examples": 0
2147
+ },
2148
+ "var_assign_d0_code_prefix_query": {
2149
+ "loss": 0.2906272601459197,
2150
+ "ppl": 1.337266038693835,
2151
+ "accuracy": 0.876,
2152
+ "loss_sum": 765.5122032243526,
2153
+ "num_tokens": 2634,
2154
+ "examples": 1000,
2155
+ "cropped_examples": 0
2156
+ },
2157
+ "var_assign_d1_math_prefix_query": {
2158
+ "loss": 2.101886375468256,
2159
+ "ppl": 8.181588913600711,
2160
+ "accuracy": 0.131,
2161
+ "loss_sum": 3476.520065024495,
2162
+ "num_tokens": 1654,
2163
+ "examples": 1000,
2164
+ "cropped_examples": 0
2165
+ },
2166
+ "var_assign_d1_code_prefix_query": {
2167
+ "loss": 0.9714342547626851,
2168
+ "ppl": 2.641730658296568,
2169
+ "accuracy": 0.312,
2170
+ "loss_sum": 2582.072249159217,
2171
+ "num_tokens": 2658,
2172
+ "examples": 1000,
2173
+ "cropped_examples": 0
2174
+ }
2175
+ },
2176
+ "macro_loss": 1.0372065951556075,
2177
+ "macro_accuracy": 0.51375,
2178
+ "original_tasks": {
2179
+ "var_assign_d0_math": {
2180
+ "loss": 0.5316129144912131,
2181
+ "ppl": 1.7016747522276936,
2182
+ "accuracy": 0.46,
2183
+ "loss_sum": 1397.078739282908,
2184
+ "num_tokens": 2628,
2185
+ "examples": 1000,
2186
+ "cropped_examples": 0
2187
+ },
2188
+ "var_assign_d0_code": {
2189
+ "loss": 0.45201226381900395,
2190
+ "ppl": 1.5714712206980943,
2191
+ "accuracy": 0.529,
2192
+ "loss_sum": 1190.6003028992563,
2193
+ "num_tokens": 2634,
2194
+ "examples": 1000,
2195
+ "cropped_examples": 0
2196
+ },
2197
+ "var_assign_d1_math": {
2198
+ "loss": 0.8022562178317106,
2199
+ "ppl": 2.230567902462197,
2200
+ "accuracy": 0.259,
2201
+ "loss_sum": 2129.18800212536,
2202
+ "num_tokens": 2654,
2203
+ "examples": 1000,
2204
+ "cropped_examples": 0
2205
+ },
2206
+ "var_assign_d1_code": {
2207
+ "loss": 0.8188754530301902,
2208
+ "ppl": 2.2679479889258873,
2209
+ "accuracy": 0.27,
2210
+ "loss_sum": 2176.5709541542456,
2211
+ "num_tokens": 2658,
2212
+ "examples": 1000,
2213
+ "cropped_examples": 0
2214
+ }
2215
+ }
2216
+ },
2217
+ {
2218
+ "name": "r3_d0_p0_a2_24000",
2219
+ "checkpoint_name": "r3_D0-P0-A2-R0_7b_seed42_20260919-160042/step_24000",
2220
+ "checkpoint": "/taijifs_wzbd31/rentTest/hunyuan/kenanchen/loopTF-workspace/output/paper2604-all-20260920/checkpoints/r3_d0_p0_a2_24000/evaluation_native.pt",
2221
+ "checkpoint_sha256": "44d07efbb7c219a7505a8084af8b799ecc14834499856efb6dd350f148338ae3",
2222
+ "step": 24000,
2223
+ "architecture": "lookahead",
2224
+ "parameter_count": 659332608,
2225
+ "training_tokens": 6291456000,
2226
+ "inference": {
2227
+ "requested": "auto",
2228
+ "resolved": "fused",
2229
+ "full_bandwidth": true,
2230
+ "prompt_traversals": 2,
2231
+ "fused_prefill_passes": 1,
2232
+ "continuation": "one feedback transition per real token",
2233
+ "cache": "architecture-owned immutable real-prefix state"
2234
+ },
2235
+ "original_result_dir": "/root/dl/runtime/script/loopTF-workspace/output/benchmarks/20260920-103547_r3_d0_p0_a2_24000-full",
2236
+ "original_policy": {
2237
+ "mode": "parallel",
2238
+ "num_passes": 3,
2239
+ "feedback_loss_weight": 1.0,
2240
+ "attention_backend": "sdpa_auto",
2241
+ "checkpoint_refiner": false,
2242
+ "cache_layer_indices": null,
2243
+ "refiner_intermediate_size": 5632,
2244
+ "refiner_gate_bias": -2.0,
2245
+ "future_loss_weight": 0.0,
2246
+ "future_gradient_mode": "all",
2247
+ "draft_loss_weight": 0.0,
2248
+ "refined_future_loss_weight": 0.0,
2249
+ "pseudo_initializer": "A2",
2250
+ "pseudo_top_k": 128,
2251
+ "pseudo_temperature": 1.0,
2252
+ "refiner_variant": "R0",
2253
+ "allow_objective_change_on_resume": false
2254
+ },
2255
+ "status": "succeeded",
2256
+ "completed_tasks": 4,
2257
+ "node": "22.169.167.168",
2258
+ "gpu": 5,
2259
+ "result_dir": "/root/dl/runtime/script/loopTF-workspace/output/benchmarks/20260920-141240_r3_d0_p0_a2_24000-prefix-full",
2260
+ "tasks": {
2261
+ "var_assign_d0_math_prefix_query": {
2262
+ "loss": 0.8231659767787711,
2263
+ "ppl": 2.2776995783865877,
2264
+ "accuracy": 0.803,
2265
+ "loss_sum": 1340.1142101958394,
2266
+ "num_tokens": 1628,
2267
+ "examples": 1000,
2268
+ "cropped_examples": 0
2269
+ },
2270
+ "var_assign_d0_code_prefix_query": {
2271
+ "loss": 0.32827057326092307,
2272
+ "ppl": 1.3885646299255863,
2273
+ "accuracy": 0.94,
2274
+ "loss_sum": 864.6646899692714,
2275
+ "num_tokens": 2634,
2276
+ "examples": 1000,
2277
+ "cropped_examples": 0
2278
+ },
2279
+ "var_assign_d1_math_prefix_query": {
2280
+ "loss": 2.4486134385615768,
2281
+ "ppl": 11.572289898983826,
2282
+ "accuracy": 0.041,
2283
+ "loss_sum": 4050.006627380848,
2284
+ "num_tokens": 1654,
2285
+ "examples": 1000,
2286
+ "cropped_examples": 0
2287
+ },
2288
+ "var_assign_d1_code_prefix_query": {
2289
+ "loss": 1.2721769482099177,
2290
+ "ppl": 3.568612797717448,
2291
+ "accuracy": 0.223,
2292
+ "loss_sum": 3381.446328341961,
2293
+ "num_tokens": 2658,
2294
+ "examples": 1000,
2295
+ "cropped_examples": 0
2296
+ }
2297
+ },
2298
+ "macro_loss": 1.2180567342027973,
2299
+ "macro_accuracy": 0.50175,
2300
+ "original_tasks": {
2301
+ "var_assign_d0_math": {
2302
+ "loss": 0.5522988319795301,
2303
+ "ppl": 1.737242058644022,
2304
+ "accuracy": 0.614,
2305
+ "loss_sum": 1451.441330442205,
2306
+ "num_tokens": 2628,
2307
+ "examples": 1000,
2308
+ "cropped_examples": 0
2309
+ },
2310
+ "var_assign_d0_code": {
2311
+ "loss": 0.5160644471499681,
2312
+ "ppl": 1.675420950023512,
2313
+ "accuracy": 0.745,
2314
+ "loss_sum": 1359.313753793016,
2315
+ "num_tokens": 2634,
2316
+ "examples": 1000,
2317
+ "cropped_examples": 0
2318
+ },
2319
+ "var_assign_d1_math": {
2320
+ "loss": 0.855579585472112,
2321
+ "ppl": 2.352737598017731,
2322
+ "accuracy": 0.244,
2323
+ "loss_sum": 2270.7082198429853,
2324
+ "num_tokens": 2654,
2325
+ "examples": 1000,
2326
+ "cropped_examples": 0
2327
+ },
2328
+ "var_assign_d1_code": {
2329
+ "loss": 0.9651471290357574,
2330
+ "ppl": 2.6251738673622613,
2331
+ "accuracy": 0.201,
2332
+ "loss_sum": 2565.361068977043,
2333
+ "num_tokens": 2658,
2334
+ "examples": 1000,
2335
+ "cropped_examples": 0
2336
+ }
2337
+ }
2338
+ },
2339
+ {
2340
+ "name": "r3_d0_p0_a2_r1_24000",
2341
+ "checkpoint_name": "r3_D0-P0-A2-R1_7b_seed42_20260919-161651/step_24000",
2342
+ "checkpoint": "/taijifs_wzbd31/rentTest/hunyuan/kenanchen/loopTF-workspace/output/paper2604-all-20260920/checkpoints/r3_d0_p0_a2_r1_24000/evaluation_native.pt",
2343
+ "checkpoint_sha256": "e4f65a8a58ed68c38e0ff5fe4ff4852753b083392923473aea0b95b2b08e80f6",
2344
+ "step": 24000,
2345
+ "architecture": "lookahead",
2346
+ "parameter_count": 665275776,
2347
+ "training_tokens": 6291456000,
2348
+ "inference": {
2349
+ "requested": "auto",
2350
+ "resolved": "fused",
2351
+ "full_bandwidth": true,
2352
+ "prompt_traversals": 2,
2353
+ "fused_prefill_passes": 1,
2354
+ "continuation": "one feedback transition per real token",
2355
+ "cache": "architecture-owned immutable real-prefix state"
2356
+ },
2357
+ "original_result_dir": "/root/dl/runtime/script/loopTF-workspace/output/benchmarks/20260920-103714_r3_d0_p0_a2_r1_24000-full",
2358
+ "original_policy": {
2359
+ "mode": "parallel",
2360
+ "num_passes": 3,
2361
+ "feedback_loss_weight": 1.0,
2362
+ "attention_backend": "sdpa_auto",
2363
+ "checkpoint_refiner": false,
2364
+ "cache_layer_indices": null,
2365
+ "refiner_intermediate_size": 5632,
2366
+ "refiner_gate_bias": -2.0,
2367
+ "future_loss_weight": 0.0,
2368
+ "future_gradient_mode": "all",
2369
+ "draft_loss_weight": 0.0,
2370
+ "refined_future_loss_weight": 1.0,
2371
+ "pseudo_initializer": "A2",
2372
+ "pseudo_top_k": 128,
2373
+ "pseudo_temperature": 1.0,
2374
+ "refiner_variant": "R1",
2375
+ "allow_objective_change_on_resume": false
2376
+ },
2377
+ "status": "succeeded",
2378
+ "completed_tasks": 4,
2379
+ "node": "22.169.130.39",
2380
+ "gpu": 0,
2381
+ "result_dir": "/root/dl/runtime/script/loopTF-workspace/output/benchmarks/20260920-141410_r3_d0_p0_a2_r1_24000-prefix-full",
2382
+ "tasks": {
2383
+ "var_assign_d0_math_prefix_query": {
2384
+ "loss": 0.5926108945742322,
2385
+ "ppl": 1.8087045930057695,
2386
+ "accuracy": 0.861,
2387
+ "loss_sum": 964.7705363668501,
2388
+ "num_tokens": 1628,
2389
+ "examples": 1000,
2390
+ "cropped_examples": 0
2391
+ },
2392
+ "var_assign_d0_code_prefix_query": {
2393
+ "loss": 0.1989325830450737,
2394
+ "ppl": 1.2200997077201048,
2395
+ "accuracy": 0.951,
2396
+ "loss_sum": 523.9884237407241,
2397
+ "num_tokens": 2634,
2398
+ "examples": 1000,
2399
+ "cropped_examples": 0
2400
+ },
2401
+ "var_assign_d1_math_prefix_query": {
2402
+ "loss": 2.0504773356642083,
2403
+ "ppl": 7.771609887637235,
2404
+ "accuracy": 0.109,
2405
+ "loss_sum": 3391.4895131886005,
2406
+ "num_tokens": 1654,
2407
+ "examples": 1000,
2408
+ "cropped_examples": 0
2409
+ },
2410
+ "var_assign_d1_code_prefix_query": {
2411
+ "loss": 1.0810439685111526,
2412
+ "ppl": 2.947755309010144,
2413
+ "accuracy": 0.274,
2414
+ "loss_sum": 2873.4148683026433,
2415
+ "num_tokens": 2658,
2416
+ "examples": 1000,
2417
+ "cropped_examples": 0
2418
+ }
2419
+ },
2420
+ "macro_loss": 0.9807661954486667,
2421
+ "macro_accuracy": 0.54875,
2422
+ "original_tasks": {
2423
+ "var_assign_d0_math": {
2424
+ "loss": 0.4579841875512992,
2425
+ "ppl": 1.5808840052165474,
2426
+ "accuracy": 0.598,
2427
+ "loss_sum": 1203.5824448848143,
2428
+ "num_tokens": 2628,
2429
+ "examples": 1000,
2430
+ "cropped_examples": 0
2431
+ },
2432
+ "var_assign_d0_code": {
2433
+ "loss": 0.32471929304337094,
2434
+ "ppl": 1.3836421934734482,
2435
+ "accuracy": 0.73,
2436
+ "loss_sum": 855.3106178762391,
2437
+ "num_tokens": 2634,
2438
+ "examples": 1000,
2439
+ "cropped_examples": 0
2440
+ },
2441
+ "var_assign_d1_math": {
2442
+ "loss": 0.8788297075382416,
2443
+ "ppl": 2.4080798989862546,
2444
+ "accuracy": 0.227,
2445
+ "loss_sum": 2332.4140438064933,
2446
+ "num_tokens": 2654,
2447
+ "examples": 1000,
2448
+ "cropped_examples": 0
2449
+ },
2450
+ "var_assign_d1_code": {
2451
+ "loss": 0.8382168725830909,
2452
+ "ppl": 2.312240279624009,
2453
+ "accuracy": 0.264,
2454
+ "loss_sum": 2227.9804473258555,
2455
+ "num_tokens": 2658,
2456
+ "examples": 1000,
2457
+ "cropped_examples": 0
2458
+ }
2459
+ }
2460
+ },
2461
+ {
2462
+ "name": "r3_d1_a0_24000",
2463
+ "checkpoint_name": "r3_D1-A0-R0_7b_seed42_20260919-040211/step_24000",
2464
+ "checkpoint": "/taijifs_wzbd31/rentTest/hunyuan/kenanchen/loopTF-workspace/output/paper2604-expanded-20260920/checkpoints/r3_d1_a0_24000/evaluation_native.pt",
2465
+ "checkpoint_sha256": "c9eef596a3d7295be4571199714ba9655c25e6d3a7687d3fb9f432f6fdc6344a",
2466
+ "step": 24000,
2467
+ "architecture": "lookahead",
2468
+ "parameter_count": 661315072,
2469
+ "training_tokens": 6291456000,
2470
+ "inference": {
2471
+ "requested": "auto",
2472
+ "resolved": "fused",
2473
+ "full_bandwidth": true,
2474
+ "prompt_traversals": 2,
2475
+ "fused_prefill_passes": 1,
2476
+ "continuation": "one feedback transition per real token",
2477
+ "cache": "architecture-owned immutable real-prefix state"
2478
+ },
2479
+ "original_result_dir": "/root/dl/runtime/script/loopTF-workspace/output/benchmarks/20260920-084133_r3_d1_a0_24000-full",
2480
+ "original_policy": {
2481
+ "mode": "parallel",
2482
+ "num_passes": 3,
2483
+ "feedback_loss_weight": 1.0,
2484
+ "attention_backend": "sdpa_auto",
2485
+ "checkpoint_refiner": false,
2486
+ "cache_layer_indices": null,
2487
+ "refiner_intermediate_size": 5632,
2488
+ "refiner_gate_bias": -2.0,
2489
+ "future_loss_weight": 0.2,
2490
+ "future_gradient_mode": "all",
2491
+ "draft_loss_weight": 0.2,
2492
+ "refined_future_loss_weight": 0.0,
2493
+ "pseudo_initializer": "A0",
2494
+ "pseudo_top_k": 128,
2495
+ "pseudo_temperature": 1.0,
2496
+ "refiner_variant": "R0",
2497
+ "allow_objective_change_on_resume": false
2498
+ },
2499
+ "status": "succeeded",
2500
+ "completed_tasks": 4,
2501
+ "node": "22.169.130.39",
2502
+ "gpu": 1,
2503
+ "result_dir": "/root/dl/runtime/script/loopTF-workspace/output/benchmarks/20260920-141509_r3_d1_a0_24000-prefix-full",
2504
+ "tasks": {
2505
+ "var_assign_d0_math_prefix_query": {
2506
+ "loss": 0.9869333440277734,
2507
+ "ppl": 2.682994023850296,
2508
+ "accuracy": 0.626,
2509
+ "loss_sum": 1606.7274840772152,
2510
+ "num_tokens": 1628,
2511
+ "examples": 1000,
2512
+ "cropped_examples": 0
2513
+ },
2514
+ "var_assign_d0_code_prefix_query": {
2515
+ "loss": 0.22263031088267635,
2516
+ "ppl": 1.249358614067259,
2517
+ "accuracy": 0.927,
2518
+ "loss_sum": 586.4082388649695,
2519
+ "num_tokens": 2634,
2520
+ "examples": 1000,
2521
+ "cropped_examples": 0
2522
+ },
2523
+ "var_assign_d1_math_prefix_query": {
2524
+ "loss": 2.457737303108085,
2525
+ "ppl": 11.678357039833326,
2526
+ "accuracy": 0.085,
2527
+ "loss_sum": 4065.0974993407726,
2528
+ "num_tokens": 1654,
2529
+ "examples": 1000,
2530
+ "cropped_examples": 0
2531
+ },
2532
+ "var_assign_d1_code_prefix_query": {
2533
+ "loss": 1.1578114104728396,
2534
+ "ppl": 3.1829594557545127,
2535
+ "accuracy": 0.256,
2536
+ "loss_sum": 3077.462729036808,
2537
+ "num_tokens": 2658,
2538
+ "examples": 1000,
2539
+ "cropped_examples": 0
2540
+ }
2541
+ },
2542
+ "macro_loss": 1.2062780921228435,
2543
+ "macro_accuracy": 0.47350000000000003,
2544
+ "original_tasks": {
2545
+ "var_assign_d0_math": {
2546
+ "loss": 0.5813156479513882,
2547
+ "ppl": 1.788389774984035,
2548
+ "accuracy": 0.452,
2549
+ "loss_sum": 1527.6975228162482,
2550
+ "num_tokens": 2628,
2551
+ "examples": 1000,
2552
+ "cropped_examples": 0
2553
+ },
2554
+ "var_assign_d0_code": {
2555
+ "loss": 0.39572687419668834,
2556
+ "ppl": 1.4854635437165922,
2557
+ "accuracy": 0.652,
2558
+ "loss_sum": 1042.3445866340771,
2559
+ "num_tokens": 2634,
2560
+ "examples": 1000,
2561
+ "cropped_examples": 0
2562
+ },
2563
+ "var_assign_d1_math": {
2564
+ "loss": 1.0979691676559669,
2565
+ "ppl": 2.9980712572375485,
2566
+ "accuracy": 0.123,
2567
+ "loss_sum": 2914.010170958936,
2568
+ "num_tokens": 2654,
2569
+ "examples": 1000,
2570
+ "cropped_examples": 0
2571
+ },
2572
+ "var_assign_d1_code": {
2573
+ "loss": 0.9949684555132847,
2574
+ "ppl": 2.7046390234840603,
2575
+ "accuracy": 0.211,
2576
+ "loss_sum": 2644.626154754311,
2577
+ "num_tokens": 2658,
2578
+ "examples": 1000,
2579
+ "cropped_examples": 0
2580
+ }
2581
+ }
2582
+ },
2583
+ {
2584
+ "name": "r3_d1_a2_24000",
2585
+ "checkpoint_name": "r3_d1_a2_r0_7b_seed42_20260919-035203-resume/step_24000",
2586
+ "checkpoint": "/taijifs_wzbd31/rentTest/hunyuan/kenanchen/loopTF-workspace/output/paper2604-expanded-20260920/checkpoints/r3_d1_a2_24000/evaluation_native.pt",
2587
+ "checkpoint_sha256": "60c796a2b809a2a7f8fcf5bf6a2cd5d52c1f29204bfab95c248d6d3f4ea68db3",
2588
+ "step": 24000,
2589
+ "architecture": "lookahead",
2590
+ "parameter_count": 659332608,
2591
+ "training_tokens": 6291456000,
2592
+ "inference": {
2593
+ "requested": "auto",
2594
+ "resolved": "fused",
2595
+ "full_bandwidth": true,
2596
+ "prompt_traversals": 2,
2597
+ "fused_prefill_passes": 1,
2598
+ "continuation": "one feedback transition per real token",
2599
+ "cache": "architecture-owned immutable real-prefix state"
2600
+ },
2601
+ "original_result_dir": "/root/dl/runtime/script/loopTF-workspace/output/benchmarks/20260920-084134_r3_d1_a2_24000-full",
2602
+ "original_policy": {
2603
+ "mode": "parallel",
2604
+ "num_passes": 3,
2605
+ "feedback_loss_weight": 1.0,
2606
+ "attention_backend": "sdpa_auto",
2607
+ "checkpoint_refiner": false,
2608
+ "cache_layer_indices": null,
2609
+ "refiner_intermediate_size": 5632,
2610
+ "refiner_gate_bias": -2.0,
2611
+ "future_loss_weight": 0.2,
2612
+ "future_gradient_mode": "all",
2613
+ "draft_loss_weight": 0.2,
2614
+ "refined_future_loss_weight": 0.0,
2615
+ "pseudo_initializer": "A2",
2616
+ "pseudo_top_k": 128,
2617
+ "pseudo_temperature": 1.0,
2618
+ "refiner_variant": "R0",
2619
+ "allow_objective_change_on_resume": false
2620
+ },
2621
+ "status": "succeeded",
2622
+ "completed_tasks": 4,
2623
+ "node": "22.169.130.39",
2624
+ "gpu": 2,
2625
+ "result_dir": "/root/dl/runtime/script/loopTF-workspace/output/benchmarks/20260920-141608_r3_d1_a2_24000-prefix-full",
2626
+ "tasks": {
2627
+ "var_assign_d0_math_prefix_query": {
2628
+ "loss": 0.9424619184353867,
2629
+ "ppl": 2.5662916482257465,
2630
+ "accuracy": 0.652,
2631
+ "loss_sum": 1534.3280032128096,
2632
+ "num_tokens": 1628,
2633
+ "examples": 1000,
2634
+ "cropped_examples": 0
2635
+ },
2636
+ "var_assign_d0_code_prefix_query": {
2637
+ "loss": 0.2954758082943171,
2638
+ "ppl": 1.3437655814097327,
2639
+ "accuracy": 0.853,
2640
+ "loss_sum": 778.2832790472312,
2641
+ "num_tokens": 2634,
2642
+ "examples": 1000,
2643
+ "cropped_examples": 0
2644
+ },
2645
+ "var_assign_d1_math_prefix_query": {
2646
+ "loss": 2.516175852144788,
2647
+ "ppl": 12.381158635132149,
2648
+ "accuracy": 0.063,
2649
+ "loss_sum": 4161.754859447479,
2650
+ "num_tokens": 1654,
2651
+ "examples": 1000,
2652
+ "cropped_examples": 0
2653
+ },
2654
+ "var_assign_d1_code_prefix_query": {
2655
+ "loss": 1.2441814895312844,
2656
+ "ppl": 3.4700933288916938,
2657
+ "accuracy": 0.166,
2658
+ "loss_sum": 3307.034399174154,
2659
+ "num_tokens": 2658,
2660
+ "examples": 1000,
2661
+ "cropped_examples": 0
2662
+ }
2663
+ },
2664
+ "macro_loss": 1.249573767101444,
2665
+ "macro_accuracy": 0.4335,
2666
+ "original_tasks": {
2667
+ "var_assign_d0_math": {
2668
+ "loss": 0.3702508469095916,
2669
+ "ppl": 1.4480978199698813,
2670
+ "accuracy": 0.715,
2671
+ "loss_sum": 973.0192256784067,
2672
+ "num_tokens": 2628,
2673
+ "examples": 1000,
2674
+ "cropped_examples": 0
2675
+ },
2676
+ "var_assign_d0_code": {
2677
+ "loss": 0.28673188804932387,
2678
+ "ppl": 1.3320670225063962,
2679
+ "accuracy": 0.785,
2680
+ "loss_sum": 755.251793121919,
2681
+ "num_tokens": 2634,
2682
+ "examples": 1000,
2683
+ "cropped_examples": 0
2684
+ },
2685
+ "var_assign_d1_math": {
2686
+ "loss": 0.9630934728468125,
2687
+ "ppl": 2.6197881948551753,
2688
+ "accuracy": 0.19,
2689
+ "loss_sum": 2556.0500769354403,
2690
+ "num_tokens": 2654,
2691
+ "examples": 1000,
2692
+ "cropped_examples": 0
2693
+ },
2694
+ "var_assign_d1_code": {
2695
+ "loss": 0.8750571668501228,
2696
+ "ppl": 2.3990124340313868,
2697
+ "accuracy": 0.261,
2698
+ "loss_sum": 2325.9019494876266,
2699
+ "num_tokens": 2658,
2700
+ "examples": 1000,
2701
+ "cropped_examples": 0
2702
+ }
2703
+ }
2704
+ }
2705
+ ],
2706
+ "note": "Custom prefix-query suite, separate from the original20-task aggregate. Same assignments/answers/source indices/five-shot selections. Demonstrations and query use the new suffix. Math answer spacing matches x=4, Code x = 4; token-level loss is sensitive to this format change."
2707
+ }