File size: 2,652 Bytes
8acf99d
 
98535ac
 
 
 
53f6ad6
 
98535ac
 
 
 
8a07fa2
 
 
 
98535ac
 
 
 
 
8acf99d
 
 
 
51638ac
 
 
 
 
 
8acf99d
51638ac
3402289
51638ac
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8acf99d
3402289
8acf99d
 
51638ac
7e45396
3402289
 
51638ac
 
 
 
 
 
 
8acf99d
 
3402289
8acf99d
3402289
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
---
license: other
dataset_info:
  features:
  - name: id
    dtype: string
  - name: task
    dtype: string
  - name: text
    dtype: string
  splits:
  - name: train
    num_bytes: 1223425889
    num_examples: 1207998
  download_size: 294631286
  dataset_size: 1223425889
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---

# synth-llava

Reverse-direction augment (v5) of the synth_llava / synth_llava2
(image+caption) training text -- 1,207,998 rows (603,999 source rows x 2
tasks). **Replaces this repo's prior single-task release** (v4,
instruction-phrasing-pool augment) -- the old content isn't used by
anything going forward. `<seed2_N>` payload tokens for the forward task are
byte-identical to the prior release; the reverse task is new.

## Why this exists

Every prior release only trained image->caption (forward). Huu's own
Discord stance (2026-08-03), in response to a "isn't caption->seed2
counter-intuitive?" question: *"seed2 is bidirectional"* -- the tokenizer's
own design intent already includes text->image conditioning (its
reconstruction path is a diffusion model, SD2.1-unclip, conditioned on the
seed2 sequence), not just image understanding. Considered and initially
set aside earlier in the same investigation for lacking a cheap,
verifiable round-trip (unlike cosmos/avclm/SNAC/agent, which have real
decoders) -- but that's a property of *evaluation*, not a reason to skip
*training* on it (the same logic applies to any text-to-image diffusion
model).

## What changed

**Tokens are unchanged** -- reuses the already-encoded `<seed2_N>` sequence
from the prior release with no re-tokenization. Adds one new task per
source row:

- **forward** (unchanged, byte-identical to v4): `<seed2>...</seed2>` + one
  of 7 instruction phrasings -> `<caption>...</caption>`
- **reverse** (NEW): one of 7 "draw/imagine/visualize" instruction
  phrasings + `<caption>...</caption>` -> `<seed2>...</seed2>`. Instruction
  variant assigned deterministically per row (`sha256(id + "_reverse") % 7`).

## Format

```
[forward]
USER: <seed2> <seed2_N> ... </seed2> <one of 7 instruction phrasings> ASSISTANT:
<think>
</think>
<caption> ... </caption>

[reverse]
USER: <one of 7 draw/imagine phrasings> <caption> ... </caption> ASSISTANT:
<think>
</think>
<seed2> <seed2_N> ... </seed2>
```

## Base

Synthetic image+caption pairs from
[mixture-vitae-backup/MixtureVitae-Backup](https://huggingface.co/mixture-vitae-backup),
tokenized with the Seed2 semantic image tokenizer (`<seed2_N>`, vocab 8192)
-- the only tokenizer in this project that accepts a standalone image with
no video/action context.