luo352 commited on
Commit
cd1d94c
·
verified ·
1 Parent(s): 762995d

Add MIREL SFT and PG-OPD training data

Browse files
Qwen35-4B/README.md ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ # Qwen3.5-4B PG-OPD data
2
+
3
+ `pg_opd.parquet` is the 2,394-row clean PG-OPD dataset used by
4
+ the 4B full-parameter recipe: 1,197 verified privilege rows and 1,197 native
5
+ fallback rows. Privileged teacher text is evidence-only and excludes GT and
6
+ support indices. See `../README.md` for the original M4/Mantis data sources.
Qwen35-4B/pg_opd.metadata.json ADDED
@@ -0,0 +1,18 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "source_p1": "P1 multitask objective pool (source annotations not redistributed)",
3
+ "stage_a_results": "Internal Stage-A artifact (not redistributed)",
4
+ "stage_a_input": "Internal Stage-A artifact (not redistributed)",
5
+ "output": "pg_opd.parquet",
6
+ "selected_rows": 2394,
7
+ "verified_evidence_rows": 1197,
8
+ "fallback_native_rows": 1197,
9
+ "requested_verified_fraction": 0.5,
10
+ "actual_verified_fraction": 0.5,
11
+ "sampling_seed": 20260910,
12
+ "teacher_input_contains_gt": false,
13
+ "teacher_input_contains_support_indices": false,
14
+ "teacher_images_empty_rows": 0,
15
+ "teacher_negative_images_empty_rows": 0,
16
+ "pure_reward_free_opd": true,
17
+ "image_path_policy": "Repository-relative paths under ../../data/images/{m4,mantis}/"
18
+ }
Qwen35-4B/pg_opd.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5f0431f9d064593a375fe79ff2dd05137b54623eb56236680477ae682276d2b2
3
+ size 1937116
Qwen35-9B/README.md ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Qwen3.5-9B PG-OPD data
2
+
3
+ `pg_opd.parquet` is the 3,000-row
4
+ 9B-specific PG-OPD dataset. It was constructed from the P1 pool by the
5
+ Qwen3.5-9B SFT-LoRA checkpoint itself: 1,500 blind-verified privilege rows are
6
+ paired with 1,500 native-control rows. This is not the 4B 2,394-row dataset.
7
+
8
+ For privilege rows, the privileged teacher receives sanitized evidence content
9
+ only. The dataset does not pass GT, an answer label, or `support=[...]` locator
10
+ indices to the teacher. The student and native teacher receive the original
11
+ prompt and images.
12
+
13
+ The 9B actor is trained with a new LoRA adapter on top of the merged SFT
14
+ checkpoint; native and privileged teachers are frozen copies of that same
15
+ checkpoint.
Qwen35-9B/pg_opd.metadata.json ADDED
@@ -0,0 +1,50 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "method": "VE-OPD_9B_SFT-LoRA-specific_StageA_verified_evidence_1to1",
3
+ "p1": "P1 multitask objective pool (source annotations not redistributed)",
4
+ "stage_a_input": "Internal Stage-A artifact (not redistributed)",
5
+ "stage_a_results": "Internal Stage-A artifact (not redistributed)",
6
+ "output": "pg_opd.parquet",
7
+ "stage_a_input_rows": 8000,
8
+ "verified_privilege_available": 1735,
9
+ "selected_privilege_rows": 1500,
10
+ "selected_normal_rows": 1500,
11
+ "selected_total_rows": 3000,
12
+ "source_counts": {
13
+ "M4-COMICS_Dialogue": 1322,
14
+ "M4-DocVQA": 16,
15
+ "M4-MIT-States_PropertyCoherence": 188,
16
+ "M4-MIT-States_StateCoherence": 262,
17
+ "M4-OCR-VQA": 246,
18
+ "Mantis-imagecode": 966
19
+ },
20
+ "role_counts": {
21
+ "verified_privilege": 1500,
22
+ "native_control": 1500
23
+ },
24
+ "privilege_source_counts": {
25
+ "M4-COMICS_Dialogue": 661,
26
+ "M4-DocVQA": 8,
27
+ "M4-MIT-States_PropertyCoherence": 94,
28
+ "M4-MIT-States_StateCoherence": 131,
29
+ "M4-OCR-VQA": 123,
30
+ "Mantis-imagecode": 483
31
+ },
32
+ "privilege_family_counts": {
33
+ "book_ocr_choice": 123,
34
+ "document_ocr_choice": 8,
35
+ "image_retrieval": 483,
36
+ "property_consistency_choice": 94,
37
+ "state_consistency_choice": 131,
38
+ "visual_sequence_choice": 661
39
+ },
40
+ "student_sees_original_prompt_and_images": true,
41
+ "native_teacher_prompt_is_original": true,
42
+ "privileged_teacher_prompt_has_sanitized_evidence_only": true,
43
+ "teacher_images_equal_original_images": true,
44
+ "teacher_negative_images_equal_original_images": true,
45
+ "teacher_input_contains_gt": false,
46
+ "teacher_input_contains_support_indices": false,
47
+ "teacher_input_contains_answer_or_synthesis": false,
48
+ "sampling_seed": 20260914,
49
+ "image_path_policy": "Repository-relative paths under ../../data/images/{m4,mantis}/"
50
+ }
Qwen35-9B/pg_opd.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ae6efa9f28deee6d1a369cbd60f39833023eb0f1a0cbfefe1acf9b1254b8aee3
3
+ size 2753533
README.md ADDED
@@ -0,0 +1,107 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # MIREL training data
2
+
3
+ The release contains processed parquet files, not the original image archives.
4
+ Image references in the parquet files are repository-relative and contain no
5
+ machine-specific usernames or absolute internal paths. Put the downloaded
6
+ image archives under `final/data/images/` (or rewrite the paths as described
7
+ below) before training.
8
+
9
+ ## SFT data
10
+
11
+ `SFT/verified_cot.parquet` contains 2,698 unique,
12
+ verified structured-CoT examples. It is deduplicated by `task_uid` from the
13
+ historical 5,396-row file; no example is intentionally repeated in this
14
+ release. The source tasks are:
15
+
16
+ - `M4-MIT-States_PropertyCoherence`: 1,054 rows;
17
+ - `M4-MIT-States_StateCoherence`: 594 rows;
18
+ - `Mantis-dreamsim`: 638 rows;
19
+ - `Mantis-imagecode`: 412 rows.
20
+
21
+ The messages include the system prompt used for CoT-SFT and set
22
+ `enable_thinking=false`.
23
+
24
+ ## PG-OPD data
25
+
26
+ The two model directories intentionally contain model-specific data; they are
27
+ not interchangeable.
28
+
29
+ ### Qwen3.5-4B
30
+
31
+ `Qwen35-4B/pg_opd.parquet` contains 2,394 rows:
32
+
33
+ - 1,197 verified evidence-privileged rows;
34
+ - 1,197 native fallback rows;
35
+ - evidence was selected from the 4B Stage-A filtering run.
36
+
37
+ ### Qwen3.5-9B
38
+
39
+ `Qwen35-9B/pg_opd.parquet` contains
40
+ 3,000 rows:
41
+
42
+ - 1,500 verified privilege rows generated and blind-verified with the 9B
43
+ SFT-LoRA checkpoint;
44
+ - 1,500 native-control rows matched from the same P1 pool;
45
+ - source Stage-A pool: 8,000 P1 samples, with 1,735 verified privilege rows
46
+ available before the deterministic 1,500-row selection.
47
+
48
+ For both recipes, privilege text contains evidence content only: no GT answer,
49
+ answer label, or `support=[...]` locator is passed to the privileged teacher.
50
+ Rows retain the native prompt/images and the teacher-side evidence fields. The
51
+ teacher advantage is computed online by comparing fixed native and privileged
52
+ teachers on the same student response tokens. There is no GRPO answer reward
53
+ in PG-OPD.
54
+
55
+ ## Original data to download
56
+
57
+ ### M4-Instruct
58
+
59
+ Download the M4-Instruct release described by the official M4-Instruct data
60
+ card and place the following archives under one local `M4-Instruct-Data`
61
+ directory:
62
+
63
+ - `MIT-States_PropertyCoherence.zip`
64
+ - `MIT-States_StateCoherence.zip`
65
+ - `COMICS_Dialogue.zip` (needed by the P1 objective-data construction)
66
+ - `DocVQA.zip`
67
+ - `OCR-VQA.zip`
68
+
69
+ The M4 data card and license are included in the upstream M4-Instruct release;
70
+ the project copy records CC BY 4.0. The processed SFT file uses the two
71
+ MIT-States subsets. The historical P1 objective construction additionally
72
+ uses the COMICS, DocVQA and OCR-VQA subsets.
73
+
74
+ ### Mantis-Instruct
75
+
76
+ Download the upstream **TIGER-Lab/Mantis-Instruct** release and its image
77
+ archives. The processed files use these Mantis subsets:
78
+
79
+ - `Mantis-dreamsim` for verified SFT;
80
+ - `Mantis-imagecode` for verified SFT and objective PG/GRPO construction.
81
+
82
+ Use the upstream dataset's `script` revision when downloading image archives,
83
+ as its dataset card explains that the images are provided separately. Follow
84
+ the upstream license and citation requirements.
85
+
86
+ Official references:
87
+
88
+ - https://github.com/TIGER-AI-Lab/Mantis
89
+ - https://huggingface.co/datasets/TIGER-Lab/Mantis-Instruct
90
+ - https://llava-vl.github.io/blog/2024-06-16-llava-next-interleave/ (M4-Instruct)
91
+
92
+ ## Image layout and path rewriting
93
+
94
+ The sanitized parquet paths use this layout relative to the `PG-OPD` working
95
+ directory:
96
+
97
+ ```text
98
+ ../../data/images/m4/...
99
+ ../../data/images/mantis/...
100
+ ```
101
+
102
+ Place the M4 and Mantis image archives below `final/data/images/m4/` and
103
+ `final/data/images/mantis/`, preserving the suffixes recorded in the parquet
104
+ files. Alternatively, use `code/PG-OPD/scripts/rewrite_image_paths.py` to map
105
+ `../../data/images` to another local image root. The helper changes only image
106
+ path fields; row order and textual fields remain unchanged. Run its
107
+ `--check-only` mode before training.
SFT/README.md ADDED
@@ -0,0 +1,11 @@
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # SFT data
2
+
3
+ `verified_cot.parquet` is the release SFT file.
4
+ It contains 2,698 unique verified structured-CoT rows, deduplicated by
5
+ `task_uid`. The historical 5,396-row file repeated each row twice; the
6
+ repetition is intentionally not included here.
7
+
8
+ The four source task groups are M4 MIT-States PropertyCoherence, M4
9
+ MIT-States StateCoherence, Mantis-dreamsim, and Mantis-imagecode. The model is
10
+ trained with the original system prompt stored in each row and with thinking
11
+ disabled.
SFT/verified_cot.manifest.json ADDED
@@ -0,0 +1,20 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "source": "Internal verified-CoT source (not redistributed)",
3
+ "output": "verified_cot.parquet",
4
+ "rows": 2698,
5
+ "source_rows": 5396,
6
+ "dedup_key": "task_uid",
7
+ "unique_task_uids": 2698,
8
+ "source_counts": {
9
+ "M4-MIT-States_PropertyCoherence": 1054,
10
+ "M4-MIT-States_StateCoherence": 594,
11
+ "Mantis-dreamsim": 638,
12
+ "Mantis-imagecode": 412
13
+ },
14
+ "training_variant": "verified_structured_cot_only",
15
+ "system_prompt_in_messages": true,
16
+ "enable_thinking_values": [
17
+ false
18
+ ],
19
+ "image_path_policy": "Repository-relative paths under ../../data/images/{m4,mantis}/"
20
+ }
SFT/verified_cot.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:261cb953f6bd2101d1d452a5623a511568f26a3be15699c848629d0ab0479c82
3
+ size 843756