Add Gemma-4-12B recaptions for all 1,000,000 images

Two captions per image from google/gemma-4-12B-it with the same prompt, bans and filter as recaptions/. Code updated to support Gemma (prh-vlm@8f8a8bf); the Qwen prompt is unchanged.

README.md

@@ -146,6 +146,44 @@ qwen3 = load_dataset(
 )
 ```
 
+## Gemma-4-12B recaptions
+
+`gemma-4-12b-recaptions/<shard>.parquet` has the same layout as `recaptions/` (same columns, same order, 1,000,000 rows), but both new captions were written by [google/gemma-4-12B-it](https://huggingface.co/google/gemma-4-12B-it). Join on `key`.
+
+```python
+from datasets import load_dataset
+
+gemma = load_dataset(
+    "parquet",
+    data_files="hf://datasets/IntelligenceLab/BLIP3o-Recap/gemma-4-12b-recaptions/*.parquet",
+    split="train",
+)
+```
+
+### Generation
+
+Same code, prompt, sampling (`n=2`, temperature 0.8, top-p 0.9, at most 160 new tokens), decode-time bans and reject/resample filter as the Qwen2.5-VL-3B recaptions above, with these differences:
+
+- `--model google/gemma-4-12B-it`; the prompt is rendered with Gemma's own chat template (thinking off), and images use Gemma's default processing (no 640×28×28 cap).
+- Gemma follows "count first" literally and started 61% of captions with "One …", so a caption-initial "One" is also banned (the captions then start with "A …").
+- 12 A100 GPUs, about 9 images/s per GPU. Almost no samples were rejected.
+
+### Statistics
+
+| | |
+| --- | ---: |
+| captions that still violate a rule | 1 (a 54-word caption of a text-only image, below the 60-word minimum) |
+| words per caption (p5 / median / p95) | 80 / 84 / 91 |
+| `bert-large-uncased` tokens incl. [CLS]/[SEP] (p5 / median / p95 / p99 / max) | 99 / 106 / 117 / 126 / 188 |
+| recap_1 vs recap_2, same image: bigram Jaccard / ROUGE-L | 0.246 / 0.522 |
+| recap_1 vs recap_2, different images: bigram Jaccard / ROUGE-L | 0.039 / 0.211 |
+| recap_1 vs original caption: bigram Jaccard / ROUGE-L | 0.060 / 0.242 |
+| Gemma recap_1 vs Qwen2.5-VL-3B recap_1, same image: ROUGE-L | 0.288 |
+
+### Known limitations
+
+A manual check of 24 captions (12 images) found the main subject and its count, colors, setting, medium and legible text correct in all of them; 5 had one wrong secondary detail (an action, a clothing item, a person count, the spelling of a name, a left/right reference). No invented objects were found. Compared with the Qwen2.5-VL-3B captions, the two Gemma captions of an image are closer to each other (39% of the content words of `recap_2` are not in `recap_1`, against 57%): they mostly state the same facts in different words, with one or two different secondary details. Captions from different models (Gemma, Qwen2.5-VL-3B, the original Qwen2.5-VL-7B caption) differ more. Openings are formulaic: about 28% of captions start with "A single".
+
 ## License
 
 Apache-2.0, same as the source dataset.

code/launch.sh

@@ -1,7 +1,9 @@
 #!/bin/bash
 # One recaption.py worker per GPU inside the recap-vllm container; workers claim shards through lock files,
 # finished shards have .done (run) / .fixed (fix) markers, so rerunning this script resumes.
-#   docker exec recap-vllm bash /workspace/prh-vlm/blip3o_recap/launch.sh 0,1,2,3,4,5,6 [out_dir] [run|fix]
+#   docker exec recap-vllm bash /workspace/prh-vlm/blip3o_recap/launch.sh 0,1,2,3,4,5,6 [out_dir] [run|fix] [extra args]
+# Several nodes may share one out_dir, so locks are kept by default; CLEAN_LOCKS=1 removes stale locks (only when no
+# worker is alive on any node).
 set -e
 GPUS=${1:-0,1,2,3,4,5,6,7}
 OUT=${2:-/workspace/prh-vlm/data/blip3o_recap/recap}
@@ -9,9 +11,9 @@ MODE=${3:-run}
 cd /workspace/prh-vlm/blip3o_recap
 if pgrep -f "^python3 recaption.py" >/dev/null; then echo "workers already running" >&2; exit 1; fi
 mkdir -p "$OUT/logs"
-rm -f "$OUT"/*.lock "$OUT"/*.fixlock  # no worker is alive, so every lock is stale
+if [ "${CLEAN_LOCKS:-0}" = 1 ]; then rm -f "$OUT"/*.lock "$OUT"/*.fixlock; fi
 for g in ${GPUS//,/ }; do
   CUDA_VISIBLE_DEVICES=$g nohup python3 recaption.py $MODE --shards ../data/blip3o_recap/shards/*.tar --out "$OUT" \
-    --gpu-mem 0.85 >> "$OUT/logs/${MODE}_gpu$g.log" 2>&1 &
+    --gpu-mem 0.85 "${@:4}" >> "$OUT/logs/${MODE}_$(hostname)_gpu$g.log" 2>&1 &
 done
 sleep 2; pgrep -f "^python3 recaption.py" | wc -l

code/recaption.py

@@ -60,10 +60,12 @@ _OPENERS = [f"{a} {b}" for a in ("The", "This", "These") for b in
 # caption-initial "The"/"This"/"These" (no leading space: a different token from mid-sentence " The"); the model
 # otherwise dodges every banned opener with "The main object in the image is ...", "The mainsubject ...", etc.
 _FIRST = ("The", "This", "These")
+# Gemma obeys "count" first and opens 61% of captions with "One ..."; banning the caption-initial form gives "A ..."
+_FIRST_GEMMA = _FIRST + ("One",)
 
 
 
-def ban_bias(tokenizer) -> dict[int, float]:
+def ban_bias(tokenizer, first: tuple[str, ...] = _FIRST) -> dict[int, float]:
     """logit_bias -100 on every single-token form of _WORDS (" vibrant", "Appear", ...).
 
     vLLM's ``bad_words`` cannot ban them: it keeps the space-prefixed form only when it tokenizes to as many tokens
@@ -72,7 +74,7 @@ def ban_bias(tokenizer) -> dict[int, float]:
     """
     forms = [p + c for x in _WORDS for c in (x, x.capitalize()) for p in ("", " ")]
     ids = [tokenizer.encode(f, add_special_tokens=False) for f in forms]
-    ids += [tokenizer.encode(f, add_special_tokens=False) for f in _FIRST]
+    ids += [tokenizer.encode(f, add_special_tokens=False) for f in first]
     return {i[0]: -100.0 for i in ids if len(i) == 1}
 
 
@@ -123,11 +125,28 @@ def violations(text: str, finished: bool = True, strict: bool = True) -> list[st
     return out
 
 
+_PROMPTS: dict[bool, str] = {}  # filled by ``engine`` from the model's own chat template
+
+
 def prompt(focus: bool) -> str:
-    return (
-        f"<|im_start|>system\n{SYSTEM}<|im_end|>\n<|im_start|>user\n<|vision_start|><|image_pad|><|vision_end|>"
-        f"{USER}{FOCUS if focus else ''}<|im_end|>\n<|im_start|>assistant\n"
-    )
+    return _PROMPTS[focus]
+
+
+def build_prompts(model: str) -> None:
+    """System + user turn with one image, rendered by the model's processor (its own image placeholder)."""
+    from transformers import AutoProcessor
+
+    proc = AutoProcessor.from_pretrained(model)
+    for focus in (False, True):
+        msgs = [
+            {"role": "system", "content": [{"type": "text", "text": SYSTEM}]},
+            {"role": "user", "content": [{"type": "image"}, {"type": "text", "text": USER + (FOCUS if focus else "")}]},
+        ]
+        text = proc.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
+        bos = proc.tokenizer.bos_token
+        if bos and text.startswith(bos) and proc.tokenizer("x").input_ids[0] == proc.tokenizer.bos_token_id:
+            text = text[len(bos) :]  # vLLM tokenizes the prompt with special tokens: no double <bos> (Gemma)
+        _PROMPTS[focus] = text
 
 
 def read_shard(path: Path):
@@ -200,6 +219,7 @@ def engine(args):
     from vllm import LLM, SamplingParams
 
     _patch_bad_words()
+    build_prompts(args.model)
 
     llm = LLM(
         model=args.model,
@@ -207,11 +227,12 @@ def engine(args):
         gpu_memory_utilization=args.gpu_mem,
         enable_prefix_caching=True,
         limit_mm_per_prompt={"image": 1},
-        mm_processor_kwargs={"max_pixels": MAX_PIXELS},
+        mm_processor_kwargs={"max_pixels": MAX_PIXELS} if "qwen" in args.model.lower() else None,
         seed=args.seed,
     )
+    first = _FIRST_GEMMA if "gemma" in args.model.lower() else _FIRST
     words = [c for x in _WORDS for c in (x, x.capitalize())]  # multi-token forms; logit_bias covers single tokens
-    ban = {"bad_words": _OPENERS + _PHRASES + words, "logit_bias": ban_bias(llm.get_tokenizer())}
+    ban = {"bad_words": _OPENERS + _PHRASES + words, "logit_bias": ban_bias(llm.get_tokenizer(), first)}
     ban = {} if args.no_ban else ban
     sp = lambda n: SamplingParams(n=n, temperature=0.8, top_p=0.9, max_tokens=160, **ban)  # noqa: E731
     return llm, sp

gemma-4-12b-recaptions/sa_000000.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e0f3872ca30e115b2df9e482feed9ee4d144aa70b848a88d60ea3c0267b5aa1c
+size 5756278

gemma-4-12b-recaptions/sa_000027.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:44cf8e28002fb6b4f6bfea4fcb0be30a1eb5bc446d70247e3db01f3e74c522aa
+size 5761332

gemma-4-12b-recaptions/sa_000055.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:086026a28fd3180082833542295cac70de37f09c5fb9e1c8fc30ff26f1665e26
+size 5767603

gemma-4-12b-recaptions/sa_000083.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:72849988c012766331e987caf264752dc4ae5eda1c5139ffccac7a6f47eadfe4
+size 5764414

gemma-4-12b-recaptions/sa_000111.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b399b897d548415be315d552910e0b93c84c6514c2d0b8cca21f932b14baec99
+size 5768006

gemma-4-12b-recaptions/sa_000138.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0ac87ad39e3ba4fc18a8112423c45b35a4e2750af9b1dcc5712fe72bd3e481db
+size 5765205

gemma-4-12b-recaptions/sa_000166.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8529e3655b1a7a5af2008ae1fdfebdddd9c69d42fa1332dcc4f83f9a38eeb08d
+size 5757762

gemma-4-12b-recaptions/sa_000194.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0e5483a4ceacf696edfecd5cd5f15aefe35a64c64605e2813d9ca494a486d47e
+size 5761507

gemma-4-12b-recaptions/sa_000222.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9b8b0fcb6888a806a8b60bc338b981d502eae77e4394974a6fea3f83ea312abc
+size 5764886

gemma-4-12b-recaptions/sa_000250.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:169ea61f2523f56107f64c2a776005e5fca69e2573ff816d300b093cc39fb6f5
+size 5755530

gemma-4-12b-recaptions/sa_000277.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eb2d0cb852dcba8b351e2c92a9cd62622d3d874b18919ecc5cf9bdaf9f7a6734
+size 5756956

gemma-4-12b-recaptions/sa_000305.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:30316cce1f81e8d731443731cbce5c7a3d915c1c6c87d95c165139fcfd6485fe
+size 5755324

gemma-4-12b-recaptions/sa_000333.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1db7d05d343a5d03af7e339a2530b97a7ea311955f0929a9bd3490ad272edef5
+size 5770197

gemma-4-12b-recaptions/sa_000361.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e4ae993733237b1488be3331e3ee3f00ead32eaf89a05c27c6d392e2f98eae1a
+size 5763999

gemma-4-12b-recaptions/sa_000388.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:031997d9187cea033e19bdb41f57146403c3f4fa0c7b129c23b4e79985f2c5f9
+size 5753431

gemma-4-12b-recaptions/sa_000416.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:458c13b67adf6c38076dc4f07cb6edf7baec7997aa00e6dc6c3a119e9b78c0ea
+size 5759352

gemma-4-12b-recaptions/sa_000444.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a71f133e655f165047065ec011ba19fcb4104d55703dffeae290dd719078d415
+size 5768084

gemma-4-12b-recaptions/sa_000472.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:df6a50afb9cb4132664ac1d4e82d96883407b5de154c66888e325b905ad5cada
+size 5763160

gemma-4-12b-recaptions/sa_000500.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2dbdba1d3bc99ebcba4f63091b17b2a0f563e373e84daca6db649f07ec813a0e
+size 5756901

gemma-4-12b-recaptions/sa_000527.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c270c271eca155b97f26f6bfb5191b09920fa27261fa80633ed1a7d8ff50a5bd
+size 5769997

gemma-4-12b-recaptions/sa_000555.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:feac9ede3e7d771377d873e187b7bcf84f4fd5432a4179cd60384430b5f3277c
+size 5751204

gemma-4-12b-recaptions/sa_000583.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a114d3f481c489d63b1f0ce7052a1cf7c58e2f738501648ec631f4ea1668dea6
+size 5759143

gemma-4-12b-recaptions/sa_000611.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:adde5e061de7380788ba7181362ff937bb79c8fdd9a1a21b09c438eed33248ee
+size 5756402

gemma-4-12b-recaptions/sa_000638.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f0360cdb960c84ed1cf0d9f6af7b6ac94b06b4743687681bd973d3f30a8fffe9
+size 5775583

gemma-4-12b-recaptions/sa_000666.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fa68f7a095f96d05b5c339f29fc76dc3af7e53a78e15ab101f95e466b2179b67
+size 5769775

gemma-4-12b-recaptions/sa_000694.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:337d4f771b20a182c552f8c0506688d211a53329480d3489280866ba37d7ed2e
+size 5751891

gemma-4-12b-recaptions/sa_000722.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9eb9f30abaa8df0d9251cc5cd7f442d655b1647cf625bcef5b591b9810ffd971
+size 5758948

gemma-4-12b-recaptions/sa_000750.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:09b579bb981e22e670ab4f14bbb0030f8c70fb37b6b43d7e6cd3208d36500e98
+size 5763364

gemma-4-12b-recaptions/sa_000777.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7722e3513f750f70cb72e6112b6f87e7a6b139d1d77d1b0bd4a8aa387830b532
+size 5756422

gemma-4-12b-recaptions/sa_000805.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a8f7025f06d2cb535525da6486fde3041ca8acf7d013db778a6d04880e59caf9
+size 5750657

gemma-4-12b-recaptions/sa_000833.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8c9abe5055beebc47e31faef6f8e5b25ae32e42b78e817def27edacadc903367
+size 5744380

gemma-4-12b-recaptions/sa_000861.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d75b946bed2fec9050f3a30afdae67595746e85cd83e86c7db543697207fc95b
+size 5756952

gemma-4-12b-recaptions/sa_000888.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:de010140d640e674f65bcb5ed217e1f609b380fd7632a4ae1f5a1cff3c2252a9
+size 5748870

gemma-4-12b-recaptions/sa_000916.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c5f8f6486b1ec781071f5c6025b7f95091eb79f127c6b25e1862cdac47c58956
+size 5760220

gemma-4-12b-recaptions/sa_000944.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1ca2c5d9e6af9bc786a54b534b2ffd4b7f2726438c0ce9b75fbfe997aa34665d
+size 5742482

gemma-4-12b-recaptions/sa_000972.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7a74796b1727423d81deb2aa787bdbd2603caa82d41eb507f2c58a36a2953532
+size 5775461

gemma-4-12b-recaptions/webdataset_JDB_1.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1b56ae6f033f9a7b8baaedfff5efb4435ab2913e0b23b06754453795463e7d11
+size 5210609

gemma-4-12b-recaptions/webdataset_JDB_120.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:19982a23b0470806605e9c46d967cd1f0801af777e96d93af5c5f81535738680
+size 5243910

gemma-4-12b-recaptions/webdataset_JDB_150.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dcc846d7837e96a6972932533e8ef3258c711aa82550a57d574043766ea44de0
+size 5068398

gemma-4-12b-recaptions/webdataset_JDB_180.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9cc5ce97c9214441ce334091dfae970d2b693d1b3ee1a1a08143fd8a95a4535e
+size 5174777

gemma-4-12b-recaptions/webdataset_JDB_210.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4fc114eba3c220b818462be332cac9e5b23cc301fc9e1551718203d0f68edddc
+size 5239088

gemma-4-12b-recaptions/webdataset_JDB_240.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:de342899f773d832fffab9de0410390b37316fd76e410e70b453f33c01225168
+size 5202558

gemma-4-12b-recaptions/webdataset_JDB_270_part.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d38a845185728812356c50816875f18ec0f0c7ffd37ef4bab89a0cf9e6f9783e
+size 4434403

gemma-4-12b-recaptions/webdataset_JDB_30.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8ae3b69b76869ac2eecff54b3511f6e4226ee9e94e597c3b899588c8d73f6c8c
+size 5174488

gemma-4-12b-recaptions/webdataset_JDB_60.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:05e5d01a63e80239c3fde337f27672ec381d79505eee80a33fc3b43ab5e48f84
+size 5126484

gemma-4-12b-recaptions/webdataset_JDB_90.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f4ba400f0de0a0b6f54e7ec25b1cd781619b9a39473208010f6bfd4e92381a89
+size 5264882

gemma-4-12b-recaptions/webdataset_shard_000.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:41436eebc4c856a08ea3d8dea616606e53338e786effdec731bd82adde623ec3
+size 4755767