This is a Cross Encoder model finetuned from Qwen/Qwen3.5-0.8B on the image_to_text and text_to_image datasets using the sentence-transformers library. It computes scores for pairs of texts, which can be used for text reranking and semantic search.

Model Details

Model Description

Model Type: Cross Encoder
Base model: Qwen/Qwen3.5-0.8B
Maximum Sequence Length: 262144 tokens
Number of Output Labels: 1 label
Supported Modalities: Text, Image, Video, Message
Training Datasets:
- image_to_text
- text_to_image

Model Sources

Documentation: Sentence Transformers Documentation
Documentation: Cross Encoder Documentation
Repository: Sentence Transformers on GitHub
Hugging Face: Cross Encoders on Hugging Face

Full Model Architecture

CrossEncoder(
  (0): Transformer({'transformer_task': 'feature-extraction', 'modality_config': {'text': {'method': 'forward', 'method_output_name': 'last_hidden_state'}, 'image': {'method': 'forward', 'method_output_name': 'last_hidden_state'}, 'video': {'method': 'forward', 'method_output_name': 'last_hidden_state'}, 'message': {'method': 'forward', 'method_output_name': 'last_hidden_state'}}, 'module_output_name': 'token_embeddings', 'message_format': 'auto', 'processing_kwargs': {'chat_template': {'add_generation_prompt': True}}, 'architecture': 'Qwen3_5Model'})
  (1): Pooling({'embedding_dimension': 1024, 'pooling_mode': 'lasttoken', 'include_prompt': True})
  (2): Dense({'in_features': 1024, 'out_features': 1, 'bias': False, 'activation_function': 'torch.nn.modules.linear.Identity', 'module_input_name': 'sentence_embedding', 'module_output_name': 'scores'})
)

Usage

Direct Usage (Sentence Transformers)

First install the Sentence Transformers library:

pip install -U sentence-transformers

Then you can load this model and run inference.

python

from sentence_transformers import CrossEncoder

# Download from the 🤗 Hub
model = CrossEncoder("tomaarsen/reranker-Qwen3.5-0.8B-doodles-image-text-to-text-pooling-dense")
# Get scores for pairs of inputs
pairs = [
    ['https://huggingface.co/tomaarsen/reranker-Qwen3.5-0.8B-doodles-image-text-to-text-pooling-dense/resolve/main/assets/image_0.jpg', 'a content character with a tan head and purple puffballs hair wearing a blue fleece, green background'],
    ['https://huggingface.co/tomaarsen/reranker-Qwen3.5-0.8B-doodles-image-text-to-text-pooling-dense/resolve/main/assets/image_0.jpg', 'a grumpy character with a green head and pink hair wearing a light blue puffer, iridescent background'],
    ['https://huggingface.co/tomaarsen/reranker-Qwen3.5-0.8B-doodles-image-text-to-text-pooling-dense/resolve/main/assets/image_0.jpg', 'a surprised character with a pale head and green mullet hair wearing a blue backpack, purple background'],
    ['https://huggingface.co/tomaarsen/reranker-Qwen3.5-0.8B-doodles-image-text-to-text-pooling-dense/resolve/main/assets/image_0.jpg', 'a default character with a med head and green brushcut hair wearing a holographic sweater, gradient 4 background'],
    ['https://huggingface.co/tomaarsen/reranker-Qwen3.5-0.8B-doodles-image-text-to-text-pooling-dense/resolve/main/assets/image_0.jpg', 'a sunglasses character with a orange head and pink hair wearing a green hoodie, grey background'],
]
scores = model.predict(pairs)
print(scores)
# [0.7607 0.4866 0.5265 0.5412 0.4776]

Evaluation

Metrics

Cross Encoder Reranking

Datasets: doodles-image-to-text-eval and doodles-text-to-image-eval
Evaluated with CrossEncoderRerankingEvaluator with these parameters:
json
1{ 2 "at_k": 10 3}

Metric	doodles-image-to-text-eval	doodles-text-to-image-eval
map	0.9933	0.9175
mrr@10	0.9933	0.9175
ndcg@10	0.995	0.9386

Training Details

Training Datasets

image_to_text

Dataset: image_to_text at a575ac6
Size: 4,500 training samples
Columns: image, text, and label
Approximate statistics based on the first 1000 samples:
image text label
type image string int
details
min: 128x128 px
mean: 128x128 px
max: 128x128 px
min: 29 tokens
mean: 33.45 tokens
max: 40 tokens
0: ~80.00%
1: ~20.00%

Samples:

image	text	label
	`a cobain glasses character with a gradient 2 head and purple puffballs hair wearing a white sweater, gradient 4 background`	`1`
	`a content character with a orange head and purple long hair wearing a striped sweater, yellow background`	`0`
	`a neutral note character with a orange head and green puffballs hair wearing a combo 2 puffer, light blue background`	`0`

Loss: BinaryCrossEntropyLoss with these parameters:
json
1{ 2 "activation_fn": "torch.nn.modules.linear.Identity", 3 "pos_weight": null 4}

text_to_image

Dataset: text_to_image at a575ac6
Size: 4,500 training samples
Columns: text, image, and label
Approximate statistics based on the first 1000 samples:
text image label
type string image int
details
min: 29 tokens
mean: 33.51 tokens
max: 40 tokens
min: 128x128 px
mean: 128x128 px
max: 128x128 px
0: ~80.00%
1: ~20.00%

Samples:

text	image	label
`a cobain glasses character with a gradient 2 head and purple puffballs hair wearing a white sweater, gradient 4 background`		`1`
`a cobain glasses character with a gradient 2 head and purple puffballs hair wearing a white sweater, gradient 4 background`		`0`
`a cobain glasses character with a gradient 2 head and purple puffballs hair wearing a white sweater, gradient 4 background`		`0`

Loss: BinaryCrossEntropyLoss with these parameters:
json
1{ 2 "activation_fn": "torch.nn.modules.linear.Identity", 3 "pos_weight": null 4}

Evaluation Datasets

image_to_text

Dataset: image_to_text at a575ac6
Size: 500 evaluation samples
Columns: image, text, and label
Approximate statistics based on the first 500 samples:
image text label
type image string int
details
min: 128x128 px
mean: 128x128 px
max: 128x128 px
min: 27 tokens
mean: 33.05 tokens
max: 38 tokens
0: ~80.00%
1: ~20.00%

Samples:

image	text	label
	`a content character with a tan head and purple puffballs hair wearing a blue fleece, green background`	`1`
	`a grumpy character with a green head and pink hair wearing a light blue puffer, iridescent background`	`0`
	`a surprised character with a pale head and green mullet hair wearing a blue backpack, purple background`	`0`

Loss: BinaryCrossEntropyLoss with these parameters:
json
1{ 2 "activation_fn": "torch.nn.modules.linear.Identity", 3 "pos_weight": null 4}

text_to_image

Dataset: text_to_image at a575ac6
Size: 500 evaluation samples
Columns: text, image, and label
Approximate statistics based on the first 500 samples:
text image label
type string image int
details
min: 27 tokens
mean: 33.21 tokens
max: 38 tokens
min: 128x128 px
mean: 128x128 px
max: 128x128 px
0: ~80.00%
1: ~20.00%

Samples:

text	image	label
`a content character with a tan head and purple puffballs hair wearing a blue fleece, green background`		`1`
`a content character with a tan head and purple puffballs hair wearing a blue fleece, green background`		`0`
`a content character with a tan head and purple puffballs hair wearing a blue fleece, green background`		`0`

Loss: BinaryCrossEntropyLoss with these parameters:
json
1{ 2 "activation_fn": "torch.nn.modules.linear.Identity", 3 "pos_weight": null 4}

Training Hyperparameters

Non-Default Hyperparameters

num_train_epochs: 1
learning_rate: 5e-06
warmup_steps: 0.1
gradient_accumulation_steps: 4
bf16: True
eval_strategy: steps
per_device_eval_batch_size: 32
prompts: {'image_to_text': "Given the image, judge whether the text matches it. Respond with 1 if they match, 0 if they don't.", 'text_to_image': "Given the text, judge whether the image matches it. Respond with 1 if they match, 0 if they don't."}

All Hyperparameters

Click to expand

per_device_train_batch_size: 8
num_train_epochs: 1
max_steps: -1
learning_rate: 5e-06
lr_scheduler_type: linear
lr_scheduler_kwargs: None
warmup_steps: 0.1
optim: adamw_torch_fused
optim_args: None
weight_decay: 0.0
adam_beta1: 0.9
adam_beta2: 0.999
adam_epsilon: 1e-08
optim_target_modules: None
gradient_accumulation_steps: 4
average_tokens_across_devices: True
max_grad_norm: 1.0
label_smoothing_factor: 0.0
bf16: True
fp16: False
bf16_full_eval: False
fp16_full_eval: False
tf32: None
gradient_checkpointing: False
gradient_checkpointing_kwargs: None
torch_compile: False
torch_compile_backend: None
torch_compile_mode: None
use_liger_kernel: False
liger_kernel_config: None
use_cache: False
neftune_noise_alpha: None
torch_empty_cache_steps: None
auto_find_batch_size: False
log_on_each_node: True
logging_nan_inf_filter: True
include_num_input_tokens_seen: no
log_level: passive
log_level_replica: warning
disable_tqdm: False
project: huggingface
trackio_space_id: trackio
eval_strategy: steps
per_device_eval_batch_size: 32
prediction_loss_only: True
eval_on_start: False
eval_do_concat_batches: True
eval_use_gather_object: False
eval_accumulation_steps: None
include_for_metrics: []
batch_eval_metrics: False
save_only_model: False
save_on_each_node: False
enable_jit_checkpoint: False
push_to_hub: False
hub_private_repo: None
hub_model_id: None
hub_strategy: every_save
hub_always_push: False
hub_revision: None
load_best_model_at_end: False
ignore_data_skip: False
restore_callback_states_from_checkpoint: False
full_determinism: False
seed: 42
data_seed: None
use_cpu: False
accelerator_config: {'split_batches': False, 'dispatch_batches': None, 'even_batches': True, 'use_seedable_sampler': True, 'non_blocking': False, 'gradient_accumulation_kwargs': None}
parallelism_config: None
dataloader_drop_last: False
dataloader_num_workers: 0
dataloader_pin_memory: True
dataloader_persistent_workers: False
dataloader_prefetch_factor: None
remove_unused_columns: True
label_names: None
train_sampling_strategy: random
length_column_name: length
ddp_find_unused_parameters: None
ddp_bucket_cap_mb: None
ddp_broadcast_buffers: False
ddp_backend: None
ddp_timeout: 1800
fsdp: []
fsdp_config: {'min_num_params': 0, 'xla': False, 'xla_fsdp_v2': False, 'xla_fsdp_grad_ckpt': False}
deepspeed: None
debug: []
skip_memory_metrics: True
do_predict: False
resume_from_checkpoint: None
warmup_ratio: None
local_rank: -1
prompts: {'image_to_text': "Given the image, judge whether the text matches it. Respond with 1 if they match, 0 if they don't.", 'text_to_image': "Given the text, judge whether the image matches it. Respond with 1 if they match, 0 if they don't."}
batch_sampler: batch_sampler
multi_dataset_batch_sampler: proportional
router_mapping: {}
learning_rate_mapping: {}

Training Logs

Epoch	Step	Training Loss	image to text loss	text to image loss	doodles-image-to-text-eval_ndcg@10	doodles-text-to-image-eval_ndcg@10
-1	-1	-	-	-	0.8248	0.4728
0.1030	29	0.2942	-	-	-	-
0.2060	58	0.2192	-	-	-	-
0.2522	71	-	0.1373	0.1687	0.9457	0.8544
0.3091	87	0.1332	-	-	-	-
0.4121	116	0.0915	-	-	-	-
0.5044	142	-	0.0587	0.0508	0.995	0.9221
0.5151	145	0.0537	-	-	-	-
0.6181	174	0.0560	-	-	-	-
0.7211	203	0.0803	-	-	-	-
0.7567	213	-	0.0408	0.0427	0.995	0.9432
0.8242	232	0.0628	-	-	-	-
0.9272	261	0.0656	-	-	-	-
-1	-1	-	-	-	0.995	0.9386

Environmental Impact

Carbon emissions were measured using CodeCarbon.

Energy Consumed: 0.290 kWh
Carbon Emitted: 0.078 kg of CO2
Hours Used: 1.025 hours

Training Hardware

On Cloud: No
GPU Model: 1 x NVIDIA GeForce RTX 3090
CPU Model: 13th Gen Intel(R) Core(TM) i7-13700K
RAM Size: 31.78 GB

Framework Versions

Python: 3.11.6
Sentence Transformers: 5.4.0.dev0
Transformers: 5.3.0
PyTorch: 2.10.0+cu128
Accelerate: 1.13.0.dev0
Datasets: 4.3.0
Tokenizers: 0.22.2

Citation

BibTeX

Sentence Transformers

bibtex

@inproceedings{reimers-2019-sentence-bert,
    title = "Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks",
    author = "Reimers, Nils and Gurevych, Iryna",
    booktitle = "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing",
    month = "11",
    year = "2019",
    publisher = "Association for Computational Linguistics",
    url = "https://arxiv.org/abs/1908.10084",
}

	image	text	label
type	image	string	int
details	min: 128x128 px mean: 128x128 px max: 128x128 px	min: 29 tokens mean: 33.45 tokens max: 40 tokens	0: ~80.00% 1: ~20.00%