asferrer

gemma-4-E2B-it-oceanguard-marine-debris

README

License: apache-2.0

Files in This Repository

Table with columns: Artefact, Purpose, Consumed by
Artefact	Purpose	Consumed by
`adapter_model.safetensors` + `adapter_config.json`	LoRA adapter (about 119 MB, `r=16`, `α=32`) on top of `google/gemma-4-E2B-it`	PEFT / Unsloth pipelines
`gemma-4-E2B-it-oceanguard-Q4_K_M.gguf`	Adapter merged into the base model and quantized to Q4_K_M (about 3.4 GB, 5.88 bits per weight; the 262k-token vocabulary dominates the size). Text decoder only, llama.cpp compatible	OceanGuard Android app via `LlamaTextEngine`, or any llama.cpp client
`tokenizer.json`, `tokenizer_config.json`, `chat_template.jinja`, `processor_config.json`	Tokenizer, chat template and processor configuration saved with the adapter	Transformers / PEFT / Unsloth

A LiteRT-LM build (.litertlm) is pending; see Deployment.

Model Description

OceanGuard AI runs Gemma 4 E2B fully on-device for open-vocabulary analysis of marine debris. This adapter aligns the model with the structured output contract that the Android application consumes: a JSON array of {box_2d, label, material} objects, where box_2d is [y_min, x_min, y_max, x_max] on a 0-1000 grid and label is one of eight snake_case classes: plastic_bottle, metal_can, fishing_net, glove, face_mask, metal_debris, plastic_debris, tire. The application maps these labels onto its ecological-impact families for the health score and report generation.

The adapter is a standard PEFT LoRA (rank 16, α = 32, no dropout) attached to the attention (q/k/v/o_proj) and MLP (gate/up/down_proj) projections. It was trained on the 4-bit quantized base model (QLoRA) with Unsloth in bf16.

Intended Use

In-scope.

Mobile and edge marine-debris detection for citizen science, NGO field operations and coastal monitoring campaigns.
Research baselines for open-vocabulary detection over specialised environmental taxonomies.
A grounding component for tool-augmented reporting pipelines (see the OceanGuard agentic tool-calling architecture).

Out-of-scope.

Autonomous decision-making (cleanup dispatch, regulatory enforcement, environmental fines). This model is a triage and reporting aid; final decisions must remain with a qualified human.
Real-time safety-critical detection (collision avoidance, navigation). Latency is of the order of seconds per image.
General-purpose object detection outside the marine-debris domain; the adapter narrows the visual prior towards marine debris and may degrade on unrelated tasks.

Training Data

Table with columns: Part, Images, Instances, Notes
Part	Images	Instances	Notes
Real TRAIN (CleanSea, Neural_Ocean, Ocean_garbage)	10,245	21,123	all images of the TRAIN partition (767 CleanSea / 3,129 Neural_Ocean / 6,349 Ocean_garbage)
Synthetic	5,000	19,500	random sample (seed 42) of a clean batch of copy-paste composites made with the authors' synthetic data generator
Total	15,245		33 % synthetic

Real images come from three public marine-debris datasets, including the CleanSea corpus introduced by the model author (IbPRIA 2022):

Table with columns: Source, License, Annotations
Source	License	Annotations
CleanSea	Non-profit research or educational use (per dataset page)	COCO boxes, mapped to 8 classes
Ocean_garbage	Roboflow Universe per-dataset terms	COCO boxes, mapped to 8 classes
Neural_Ocean	Roboflow Universe per-dataset terms	COCO boxes, mapped to 8 classes

The corpus is partitioned by group (for example, all images of a CleanSea dive, or sharing a source-file name key, go to the same partition), so no group of the evaluation subsets appears in training. Training instances per class: Bottle 4,333, Can 412, Fishing_Net 2,812, Glove 2,582, Mask 3,164, Metal_Debris 102, Plastic_Debris 3,957, Tire 3,761 (real images only).

Geographic coverage. The three sources were collected predominantly in coastal waters around Japan and the wider north-western Pacific, so the training distribution is biased towards the debris types, water clarity and lighting of that region. Validate on a local hold-out before operational use elsewhere.

Re-distributors of the training corpus must comply with the original dataset licenses.

Training Procedure

Unsloth FastVisionModel, one workstation GPU, bf16 compute over a 4-bit base model. LoRA is attached to the attention and MLP projections selected by module name, which matches both the language tower and the vision encoder: the saved adapter has non-zero low-rank updates in both (245 pairs in the language tower, 112 in the vision encoder). The vision_lora switch of the training configuration had no effect on which modules were adapted (see Limitations).

Table with columns: Hyper-parameter, Value
Hyper-parameter	Value
Base model	`unsloth/gemma-4-E2B-it` (4-bit load)
LoRA `r` / `α` / dropout	16 / 32 / 0.0
Target modules	`q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`, `up_proj`, `down_proj`
Optimiser	`adamw_8bit`, learning rate 1e-4, cosine schedule, weight decay 0.01, warm-up ratio 0.1, max grad norm 1.0
Batch size	4 per device × 8 gradient accumulation = 32 effective
Epochs / optimisation steps	3 / 1,431
Sequence length / image token budget	1,024 / 64
Seed	42
Loss	causal-LM over the assistant JSON answer (`box_2d` + `label` + `material`)

Mean training loss over the run: 0.851 (the synthetic images raise it relative to a real-only run). Wall-clock 12,287 s (3.4 h) on a single NVIDIA RTX 5090 (32 GiB), Windows 11, CUDA 12.8; peak allocated GPU memory 20.6 GB.

Software: unsloth 2026.5.2, transformers 5.5.0, peft 0.19.1, torch 2.12 (nightly, CUDA 12.8), bitsandbytes. The adapter file holds 786 LoRA tensors (393 A/B pairs, about 31 M parameters; see Limitations for their distribution across towers).

Evaluation

Protocol

Data. Two stratified subsets of the group-aware partition, drawn with seed 42: TEST-200 (200 images, 433 instances: 28 CleanSea / 64 Neural_Ocean / 108 Ocean_garbage; reported results) and VAL-100 (100 images, 218 instances: 23 / 33 / 44; used to choose among configurations). Stratification is by the rarest class present in each image, with a minimum of 10 images per label. This deliberately over-represents rare classes: Metal_Debris is 3.5 % of the instances of TEST-200 against 0.5 % of the full TEST partition (2,207 images), so global figures are not estimates of full-partition performance. Annotations and download instructions are in asferrer/oceanguard-marine-debris-eval-1000 (annotations only, no images).
Inference. Same system and user prompt for all models (shown below), greedy decoding, at most 256 new tokens, JSON answers parsed into boxes; unparseable answers count as no detections.
Metric. mAP@0.5 with 11-point interpolated AP per class, averaged over the 8 classes, with every detection at score 1.0 (the model emits no confidences). A class with no prediction scores 0. A COCO-style evaluator (101-point AP, pycocotools) is also reported.
Intervals. 95 % basic bootstrap intervals over images (1,000 resamples, seed 42, the same resamples for all configurations so differences are paired).
Latency. Mean seconds per image on the RTX 5090 while other heavy processes were running on the machine; indicative only.

Configurations compared

Table with columns: Configuration, Adapted modules, Training data, Steps
Configuration	Adapted modules	Training data	Steps
Base	none	none	none
LoRA, real only (`exp10g`)	language + vision attention/MLP	10,245 real	963
LoRA, real only, second run (`exp12g`)	same as `exp10g`	10,245 real	963
LoRA, real + synthetic (`exp11g`) — this repository	same as `exp10g`	10,245 real + 5,000 synthetic	1,431

exp12g was launched with vision_lora: true, but its adapter has exactly the same structure as exp10g (the switch had no effect), so it is a second run of the real-only configuration with the same seed; the gap between the two runs is an indication of run-to-run variability.

Aggregate

Table with columns: Configuration, TEST-200 mAP@0.5, 95 % interval, VAL-100 mAP@0.5, JSON valid (TEST), Predictions (TEST), Latency (s, TEST)
Configuration	TEST-200 mAP@0.5	95 % interval	VAL-100 mAP@0.5	JSON valid (TEST)	Predictions (TEST)	Latency (s, TEST)
Base `gemma-4-E2B-it`	0.1381	[0.106, 0.187]	0.1081	0.990	183	3.19
LoRA, real only (`exp10g`)	0.2948	[0.227, 0.335]	0.2570	0.935	300	9.03
LoRA, real only, second run (`exp12g`)	0.2825	[0.211, 0.319]	0.2063	0.975	342	7.75
LoRA, real + synthetic (`exp11g`, this repo)	0.2941	[0.209, 0.320]	0.2820	0.980	313	7.04

With the COCO-style evaluator the TEST-200 mAP@0.5 is 0.1073 (base), 0.2780, 0.2704 and 0.3003 for the three LoRA rows in the order above (VAL-100: 0.0942, 0.2461, 0.2017, 0.2792).

Paired differences on TEST-200 (95 % basic bootstrap interval):

Table with columns: Comparison, Difference in mAP@0.5
Comparison	Difference in mAP@0.5
`exp10g` − base	+0.157 [0.075, 0.190]
`exp12g` − base	+0.144 [0.056, 0.177]
`exp11g` − base (this repo)	+0.156 [0.045, 0.178]
`exp11g` − `exp10g`	−0.001 [−0.072, 0.026]
`exp12g` − `exp10g`	−0.012 [−0.054, 0.021]
`exp12g` − `exp11g`	−0.012 [−0.043, 0.049]

All three adapters improve clearly over the base model, and the intervals of the differences between adapters include zero on TEST-200: adding synthetic data did not change the global mAP, and the two runs of the real-only configuration differ by 0.012 on TEST-200 but by 0.051 on VAL-100. exp11g was chosen because it had the best mAP on VAL-100 (0.2820 against 0.2570 and 0.2063), although the spread between the two real-only runs on VAL-100 shows that this choice is within run-to-run variability; it also has the highest JSON validity and the lowest (indicative) latency among the adapters.

Per-class mAP@0.5, TEST-200

n is the number of ground-truth instances of the class in TEST-200.

Table with columns: Class, n, Base, exp10g, exp12g, exp11g (this repo), Δ (exp11g − base)
Class	n	Base	`exp10g`	`exp12g`	`exp11g` (this repo)	Δ (`exp11g` − base)
Bottle	98	0.190	0.411	0.353	0.393	+0.203
Can	35	0.000	0.045	0.061	0.091	+0.091
Fishing_Net	52	0.294	0.436	0.377	0.406	+0.113
Glove	63	0.118	0.426	0.511	0.451	+0.334
Mask	67	0.091	0.251	0.250	0.343	+0.253
Metal_Debris	15	0.227	0.091	0.091	0.000	−0.227
Plastic_Debris	58	0.139	0.488	0.453	0.491	+0.352
Tire	45	0.045	0.210	0.164	0.177	+0.132

The adapter improves seven of the eight classes. Metal_Debris (15 instances in TEST-200, 102 in training) is the only class where the base model is better, and the released adapter does not detect it (0.000, the other adapters reach 0.091). Can (35 instances) is still weak. Per-class figures on so few instances are very noisy; the 5,000 synthetic images (2,041 Metal_Debris and 2,344 Can instances) did not rescue these two classes.

Per-source mAP@0.5, TEST-200

Table with columns: Source, n images, Base, exp10g, exp12g, exp11g (this repo), exp11g / base
Source	n images	Base	`exp10g`	`exp12g`	`exp11g` (this repo)	`exp11g` / base
Global	200	0.138	0.295	0.282	0.294	2.13x
CleanSea	28	0.163	0.260	0.202	0.177	1.09x
Neural_Ocean	64	0.089	0.321	0.278	0.357	4.01x
Ocean_garbage	108	0.121	0.299	0.291	0.346	2.85x

The gain concentrates in Neural_Ocean and Ocean_garbage (Roboflow-exported imagery); on CleanSea (28 images) the released adapter improves only by +0.014. The per-source and per-class tables can be recomputed from the cached predictions with scripts/compute_per_source_map.py in the evaluation dataset repository.

Prompt used for detection

System: You are a marine debris detection system. Output ONLY valid JSON arrays. Never include explanations, markdown, or text outside the JSON.

User (image + text):

text

Detect ALL marine debris and litter in this image. Output ONLY a JSON array. For each object, provide:- "box_2d": bounding box as [y_min, x_min, y_max, x_max], integers 0-1000- "label": ONE of these 8 labels (snake_case, lowercase): plastic_bottle, metal_can, fishing_net, glove, face_mask, metal_debris, plastic_debris, tire- "material": Plastic, Metal, Glass, Rubber, Fabric, Latex Output format:[{"box_2d": [y_min, x_min, y_max, x_max], "label": "object_type", "material": "material_type"}] If no debris is found, output: []

Deployment

Table with columns: Step, Tooling, Status
Step	Tooling	Status
Train LoRA	Unsloth FastVisionModel	Done
Publish adapter	Hugging Face Hub	Done (this repository)
Use adapter via Transformers / PEFT / Unsloth	`model.load_adapter(...)`	Done
Merge adapter into the base model (float16)	manual `W + (α/r)·B·A` merge	Done
Convert merged model to GGUF f16 and quantize to Q4_K_M	llama.cpp `convert_hf_to_gguf.py`, `llama-quantize`	Done: `gemma-4-E2B-it-oceanguard-Q4_K_M.gguf`
Wire the GGUF to the Android app	`LlamaTextEngine` + llama.cpp	Done (downloaded from this repository by the app)
Export merged model to LiteRT-LM (`.litertlm`)	MediaPipe converter / `ai-edge-torch`	Pending
On-device latency of the adapted model	Android test device	Pending

The GGUF contains the merged text decoder only (no vision encoder or multimodal projector). To send images through llama.cpp you need a multimodal projector (mmproj) for Gemma 4 E2B, which is not part of this repository.

GGUF download

bash

curl -L -o gemma-4-E2B-it-oceanguard-Q4_K_M.gguf \  "https://huggingface.co/asferrer/gemma-4-E2B-it-oceanguard-marine-debris/resolve/main/gemma-4-E2B-it-oceanguard-Q4_K_M.gguf" ./llama-cli -m gemma-4-E2B-it-oceanguard-Q4_K_M.gguf -n 256 --temp 0.0 -p "..."

Limitations

Small evaluation subsets. TEST-200 has 200 images and VAL-100 has 100; rare classes are over-represented by design. Intervals are wide (about ±0.06 around 0.29) and per-class and per-source numbers, especially Metal_Debris (15 instances), Can (35) and CleanSea (28 images), are noisy.
No measurable benefit from synthetic data in global TEST-200 mAP; the choice of this adapter rests on VAL-100, within run-to-run variability.
Metal_Debris is not detected (mAP 0.000) and Can is weak (0.091). Operators should expect silent misses on these classes.
JSON validity is 98 % on TEST-200 (93.5 % to 97.5 % for the other adapters): malformed answers are dropped by the parser and yield an empty detection set.
Both towers are adapted. Target modules were selected by name, so the adapter holds low-rank updates for the language tower and for same-named projections of the vision encoder (112 pairs, about 4.5 M parameters, non-zero), plus all-zero updates for the audio encoder (36 pairs, no effect). Loading the adapter with PEFT applies all of them; the GGUF contains only the language-tower part, so its outputs can differ from the PEFT adapter. The separate contribution of the vision-encoder updates has not been ablated.
Geographic and capture bias. Training imagery is predominantly from Japan and the north-western Pacific; performance elsewhere is not measured. Synthetic composites can inject lighting biases.
Fixed taxonomy. Eight classes; other debris types are mapped to the closest class or missed.
Latency figures were measured with other heavy processes on the machine and are indicative.
Not certified. Not for legal, regulatory or environmental-enforcement reporting. The OceanGuard app produces draft reports that a human reviews.
Function-calling untouched. The fine-tune targets the detection output format and is not meant to change tool-calling behaviour.

Bias, Risks and Safety

Spurious correlations. Debris classes co-occur with characteristic backgrounds (for example nets with rocky shorelines); the model may exploit context cues that do not generalise across geographies.
Sensitive categories. Medical waste is rare and high-stakes; positive detections should trigger human handling protocols rather than autonomous action.
Open-vocabulary misuse. The model can be prompted to label arbitrary objects; downstream applications should constrain inputs to the marine-debris domain and filter implausible outputs.
Privacy. In the OceanGuard app, photos are analysed on-device and never leave the phone; the adapter itself processes no user data. Outside the app, operators are responsible for compliance with applicable privacy regulation.

How to Use

A. PEFT + Unsloth

python

from unsloth import FastVisionModel model, tokenizer = FastVisionModel.from_pretrained(    "unsloth/gemma-4-E2B-it",    load_in_4bit=True,)model.load_adapter(    "asferrer/gemma-4-E2B-it-oceanguard-marine-debris",    adapter_name="oceanguard",)model.set_adapter("oceanguard")FastVisionModel.for_inference(model)

Use the detection prompt shown in Evaluation and greedy decoding (temperature = 0, max_new_tokens = 256), then parse the answer as a JSON array. An end-to-end example is in the OceanGuard repository.

B. llama.cpp (GGUF Q4_K_M)

See GGUF download. The OceanGuard Android app downloads and loads this file through its "Gemma 4 FT (GGUF)" provider using LlamaTextEngine (llama.cpp).

License and Citation

LoRA adapter: Apache 2.0.
Base model: google/gemma-4-E2B-it, subject to the Gemma Terms of Use. The merged GGUF is a derivative of the base model and is also subject to those terms.
Training data: source-dataset licenses apply (see the table in Training Data).

If you use this adapter, please cite the adapter, the application and the article:

bibtex

@misc{sanchezferrer2026oceanguard_adapter,  author       = {S{\'a}nchez-Ferrer, Alejandro},  title        = {{Gemma 4 E2B OceanGuard Marine Debris LoRA Adapter}},  year         = {2026},  howpublished = {Hugging Face Model Hub},  note         = {LoRA adapter for google/gemma-4-E2B-it. Apache 2.0.},  url          = {https://huggingface.co/asferrer/gemma-4-E2B-it-oceanguard-marine-debris}} @misc{sanchezferrer2026oceanguard_app,  author       = {S{\'a}nchez-Ferrer, Alejandro},  title        = {{OceanGuard AI: Fully Offline Marine Debris Intelligence with Gemma 4 and Two-Phase Tool Calling}},  year         = {2026},  howpublished = {Kaggle Gemma 4 Good Hackathon (Global Resilience track)},  url          = {https://github.com/asferrer/OceanguardAI-App}} @misc{sanchezferrer2026synthetic,  author       = {S{\'a}nchez-Ferrer, Alejandro and Gallego, Antonio Javier and Calvo-Zaragoza, Jorge},  title        = {{When Does Synthetic Data Help Underwater Debris Detection? Gap Metrics, Data Regimes and Detector Paradigms}},  year         = {2026},  howpublished = {Neurocomputing},  note         = {Manuscript in preparation}}

Sánchez-Ferrer, A., Valero-Mas, J. J., Gallego, A. J., & Calvo-Zaragoza, J. (2023). An experimental study on marine debris location and recognition using object detection. Pattern Recognition Letters, 168, 154-161. doi.org/10.1016/j.patrec.2022.12.019
Sánchez-Ferrer, A., Gallego, A. J., Valero-Mas, J. J., & Calvo-Zaragoza, J. (2022). The CleanSea Set: A Benchmark Corpus for Underwater Debris Detection and Recognition. IbPRIA 2022, Lecture Notes in Computer Science, Springer. doi.org/10.1007/978-3-031-04881-4_49
Sánchez-Ferrer, A. (2024). Modelos de difusión aplicados a la detección de objetos en el fondo marino. Master's Thesis, Universidad de Alicante. rua.ua.es
Sánchez-Ferrer, A. (2021). Deep Learning aplicado a la detección de residuos en el fondo marino. Bachelor's Thesis, Universidad de Alicante. rua.ua.es

Acknowledgments

Google DeepMind for releasing Gemma 4 under terms that enable open research and on-device deployment.
The Unsloth team for the FastVisionModel implementation that makes a Gemma 4 vision-language fine-tune tractable on a single workstation GPU.
CleanSea, Ocean_garbage and Neural_Ocean dataset authors for releasing marine-debris annotations.
Pattern Recognition and Artificial Intelligence (PRAI) group, Universidad de Alicante, which hosts the doctoral programme under which this work is carried out and co-authored the cited publications.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

asferrer

Model Tree

Base

google/gemma-4-E2B-it

Adapter

this model

Input Modalities

TextImage

Output Modalities

Text

Supported Functionality

Dedicated Endpoints

Explore FriendliAI today

Get started Talk to an engineer

README

License: apache-2.0

Files in This Repository

Table with columns: Artefact, Purpose, Consumed by
Artefact	Purpose	Consumed by
`adapter_model.safetensors` + `adapter_config.json`	LoRA adapter (about 119 MB, `r=16`, `α=32`) on top of `google/gemma-4-E2B-it`	PEFT / Unsloth pipelines
`gemma-4-E2B-it-oceanguard-Q4_K_M.gguf`	Adapter merged into the base model and quantized to Q4_K_M (about 3.4 GB, 5.88 bits per weight; the 262k-token vocabulary dominates the size). Text decoder only, llama.cpp compatible	OceanGuard Android app via `LlamaTextEngine`, or any llama.cpp client
`tokenizer.json`, `tokenizer_config.json`, `chat_template.jinja`, `processor_config.json`	Tokenizer, chat template and processor configuration saved with the adapter	Transformers / PEFT / Unsloth

A LiteRT-LM build (.litertlm) is pending; see Deployment.

Model Description

Intended Use

In-scope.

Mobile and edge marine-debris detection for citizen science, NGO field operations and coastal monitoring campaigns.
Research baselines for open-vocabulary detection over specialised environmental taxonomies.
A grounding component for tool-augmented reporting pipelines (see the OceanGuard agentic tool-calling architecture).

Out-of-scope.

Autonomous decision-making (cleanup dispatch, regulatory enforcement, environmental fines). This model is a triage and reporting aid; final decisions must remain with a qualified human.
Real-time safety-critical detection (collision avoidance, navigation). Latency is of the order of seconds per image.
General-purpose object detection outside the marine-debris domain; the adapter narrows the visual prior towards marine debris and may degrade on unrelated tasks.

Training Data

Table with columns: Part, Images, Instances, Notes
Part	Images	Instances	Notes
Real TRAIN (CleanSea, Neural_Ocean, Ocean_garbage)	10,245	21,123	all images of the TRAIN partition (767 CleanSea / 3,129 Neural_Ocean / 6,349 Ocean_garbage)
Synthetic	5,000	19,500	random sample (seed 42) of a clean batch of copy-paste composites made with the authors' synthetic data generator
Total	15,245		33 % synthetic

Real images come from three public marine-debris datasets, including the CleanSea corpus introduced by the model author (IbPRIA 2022):

Table with columns: Source, License, Annotations
Source	License	Annotations
CleanSea	Non-profit research or educational use (per dataset page)	COCO boxes, mapped to 8 classes
Ocean_garbage	Roboflow Universe per-dataset terms	COCO boxes, mapped to 8 classes
Neural_Ocean	Roboflow Universe per-dataset terms	COCO boxes, mapped to 8 classes

Re-distributors of the training corpus must comply with the original dataset licenses.

Training Procedure

Table with columns: Hyper-parameter, Value
Hyper-parameter	Value
Base model	`unsloth/gemma-4-E2B-it` (4-bit load)
LoRA `r` / `α` / dropout	16 / 32 / 0.0
Target modules	`q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`, `up_proj`, `down_proj`
Optimiser	`adamw_8bit`, learning rate 1e-4, cosine schedule, weight decay 0.01, warm-up ratio 0.1, max grad norm 1.0
Batch size	4 per device × 8 gradient accumulation = 32 effective
Epochs / optimisation steps	3 / 1,431
Sequence length / image token budget	1,024 / 64
Seed	42
Loss	causal-LM over the assistant JSON answer (`box_2d` + `label` + `material`)

Evaluation

Protocol

Data. Two stratified subsets of the group-aware partition, drawn with seed 42: TEST-200 (200 images, 433 instances: 28 CleanSea / 64 Neural_Ocean / 108 Ocean_garbage; reported results) and VAL-100 (100 images, 218 instances: 23 / 33 / 44; used to choose among configurations). Stratification is by the rarest class present in each image, with a minimum of 10 images per label. This deliberately over-represents rare classes: Metal_Debris is 3.5 % of the instances of TEST-200 against 0.5 % of the full TEST partition (2,207 images), so global figures are not estimates of full-partition performance. Annotations and download instructions are in asferrer/oceanguard-marine-debris-eval-1000 (annotations only, no images).
Inference. Same system and user prompt for all models (shown below), greedy decoding, at most 256 new tokens, JSON answers parsed into boxes; unparseable answers count as no detections.
Metric. mAP@0.5 with 11-point interpolated AP per class, averaged over the 8 classes, with every detection at score 1.0 (the model emits no confidences). A class with no prediction scores 0. A COCO-style evaluator (101-point AP, pycocotools) is also reported.
Intervals. 95 % basic bootstrap intervals over images (1,000 resamples, seed 42, the same resamples for all configurations so differences are paired).
Latency. Mean seconds per image on the RTX 5090 while other heavy processes were running on the machine; indicative only.

Configurations compared

Table with columns: Configuration, Adapted modules, Training data, Steps
Configuration	Adapted modules	Training data	Steps
Base	none	none	none
LoRA, real only (`exp10g`)	language + vision attention/MLP	10,245 real	963
LoRA, real only, second run (`exp12g`)	same as `exp10g`	10,245 real	963
LoRA, real + synthetic (`exp11g`) — this repository	same as `exp10g`	10,245 real + 5,000 synthetic	1,431

Aggregate

Table with columns: Configuration, TEST-200 mAP@0.5, 95 % interval, VAL-100 mAP@0.5, JSON valid (TEST), Predictions (TEST), Latency (s, TEST)
Configuration	TEST-200 mAP@0.5	95 % interval	VAL-100 mAP@0.5	JSON valid (TEST)	Predictions (TEST)	Latency (s, TEST)
Base `gemma-4-E2B-it`	0.1381	[0.106, 0.187]	0.1081	0.990	183	3.19
LoRA, real only (`exp10g`)	0.2948	[0.227, 0.335]	0.2570	0.935	300	9.03
LoRA, real only, second run (`exp12g`)	0.2825	[0.211, 0.319]	0.2063	0.975	342	7.75
LoRA, real + synthetic (`exp11g`, this repo)	0.2941	[0.209, 0.320]	0.2820	0.980	313	7.04

With the COCO-style evaluator the TEST-200 mAP@0.5 is 0.1073 (base), 0.2780, 0.2704 and 0.3003 for the three LoRA rows in the order above (VAL-100: 0.0942, 0.2461, 0.2017, 0.2792).

Paired differences on TEST-200 (95 % basic bootstrap interval):

Table with columns: Comparison, Difference in mAP@0.5
Comparison	Difference in mAP@0.5
`exp10g` − base	+0.157 [0.075, 0.190]
`exp12g` − base	+0.144 [0.056, 0.177]
`exp11g` − base (this repo)	+0.156 [0.045, 0.178]
`exp11g` − `exp10g`	−0.001 [−0.072, 0.026]
`exp12g` − `exp10g`	−0.012 [−0.054, 0.021]
`exp12g` − `exp11g`	−0.012 [−0.043, 0.049]

Per-class mAP@0.5, TEST-200

n is the number of ground-truth instances of the class in TEST-200.

Table with columns: Class, n, Base, exp10g, exp12g, exp11g (this repo), Δ (exp11g − base)
Class	n	Base	`exp10g`	`exp12g`	`exp11g` (this repo)	Δ (`exp11g` − base)
Bottle	98	0.190	0.411	0.353	0.393	+0.203
Can	35	0.000	0.045	0.061	0.091	+0.091
Fishing_Net	52	0.294	0.436	0.377	0.406	+0.113
Glove	63	0.118	0.426	0.511	0.451	+0.334
Mask	67	0.091	0.251	0.250	0.343	+0.253
Metal_Debris	15	0.227	0.091	0.091	0.000	−0.227
Plastic_Debris	58	0.139	0.488	0.453	0.491	+0.352
Tire	45	0.045	0.210	0.164	0.177	+0.132

Per-source mAP@0.5, TEST-200

Table with columns: Source, n images, Base, exp10g, exp12g, exp11g (this repo), exp11g / base
Source	n images	Base	`exp10g`	`exp12g`	`exp11g` (this repo)	`exp11g` / base
Global	200	0.138	0.295	0.282	0.294	2.13x
CleanSea	28	0.163	0.260	0.202	0.177	1.09x
Neural_Ocean	64	0.089	0.321	0.278	0.357	4.01x
Ocean_garbage	108	0.121	0.299	0.291	0.346	2.85x

Prompt used for detection

System: You are a marine debris detection system. Output ONLY valid JSON arrays. Never include explanations, markdown, or text outside the JSON.

User (image + text):

text

Detect ALL marine debris and litter in this image. Output ONLY a JSON array.For each object, provide:- "box_2d": bounding box as [y_min, x_min, y_max, x_max], integers 0-1000- "label": ONE of these 8 labels (snake_case, lowercase): plastic_bottle, metal_can, fishing_net, glove, face_mask, metal_debris, plastic_debris, tire- "material": Plastic, Metal, Glass, Rubber, Fabric, LatexOutput format:[{"box_2d": [y_min, x_min, y_max, x_max], "label": "object_type", "material": "material_type"}]If no debris is found, output: []

Detect ALL marine debris and litter in this image. Output ONLY a JSON array. For each object, provide:- "box_2d": bounding box as [y_min, x_min, y_max, x_max], integers 0-1000- "label": ONE of these 8 labels (snake_case, lowercase): plastic_bottle, metal_can, fishing_net, glove, face_mask, metal_debris, plastic_debris, tire- "material": Plastic, Metal, Glass, Rubber, Fabric, Latex Output format:[{"box_2d": [y_min, x_min, y_max, x_max], "label": "object_type", "material": "material_type"}] If no debris is found, output: []

Deployment

Table with columns: Step, Tooling, Status
Step	Tooling	Status
Train LoRA	Unsloth FastVisionModel	Done
Publish adapter	Hugging Face Hub	Done (this repository)
Use adapter via Transformers / PEFT / Unsloth	`model.load_adapter(...)`	Done
Merge adapter into the base model (float16)	manual `W + (α/r)·B·A` merge	Done
Convert merged model to GGUF f16 and quantize to Q4_K_M	llama.cpp `convert_hf_to_gguf.py`, `llama-quantize`	Done: `gemma-4-E2B-it-oceanguard-Q4_K_M.gguf`
Wire the GGUF to the Android app	`LlamaTextEngine` + llama.cpp	Done (downloaded from this repository by the app)
Export merged model to LiteRT-LM (`.litertlm`)	MediaPipe converter / `ai-edge-torch`	Pending
On-device latency of the adapted model	Android test device	Pending

GGUF download

bash

curl -L -o gemma-4-E2B-it-oceanguard-Q4_K_M.gguf \  "https://huggingface.co/asferrer/gemma-4-E2B-it-oceanguard-marine-debris/resolve/main/gemma-4-E2B-it-oceanguard-Q4_K_M.gguf" ./llama-cli -m gemma-4-E2B-it-oceanguard-Q4_K_M.gguf -n 256 --temp 0.0 -p "..."

curl -L -o gemma-4-E2B-it-oceanguard-Q4_K_M.gguf \  "https://huggingface.co/asferrer/gemma-4-E2B-it-oceanguard-marine-debris/resolve/main/gemma-4-E2B-it-oceanguard-Q4_K_M.gguf" ./llama-cli -m gemma-4-E2B-it-oceanguard-Q4_K_M.gguf -n 256 --temp 0.0 -p "..."

Limitations

Small evaluation subsets. TEST-200 has 200 images and VAL-100 has 100; rare classes are over-represented by design. Intervals are wide (about ±0.06 around 0.29) and per-class and per-source numbers, especially Metal_Debris (15 instances), Can (35) and CleanSea (28 images), are noisy.
No measurable benefit from synthetic data in global TEST-200 mAP; the choice of this adapter rests on VAL-100, within run-to-run variability.
Metal_Debris is not detected (mAP 0.000) and Can is weak (0.091). Operators should expect silent misses on these classes.
JSON validity is 98 % on TEST-200 (93.5 % to 97.5 % for the other adapters): malformed answers are dropped by the parser and yield an empty detection set.
Both towers are adapted. Target modules were selected by name, so the adapter holds low-rank updates for the language tower and for same-named projections of the vision encoder (112 pairs, about 4.5 M parameters, non-zero), plus all-zero updates for the audio encoder (36 pairs, no effect). Loading the adapter with PEFT applies all of them; the GGUF contains only the language-tower part, so its outputs can differ from the PEFT adapter. The separate contribution of the vision-encoder updates has not been ablated.
Geographic and capture bias. Training imagery is predominantly from Japan and the north-western Pacific; performance elsewhere is not measured. Synthetic composites can inject lighting biases.
Fixed taxonomy. Eight classes; other debris types are mapped to the closest class or missed.
Latency figures were measured with other heavy processes on the machine and are indicative.
Not certified. Not for legal, regulatory or environmental-enforcement reporting. The OceanGuard app produces draft reports that a human reviews.
Function-calling untouched. The fine-tune targets the detection output format and is not meant to change tool-calling behaviour.

Bias, Risks and Safety

Spurious correlations. Debris classes co-occur with characteristic backgrounds (for example nets with rocky shorelines); the model may exploit context cues that do not generalise across geographies.
Sensitive categories. Medical waste is rare and high-stakes; positive detections should trigger human handling protocols rather than autonomous action.
Open-vocabulary misuse. The model can be prompted to label arbitrary objects; downstream applications should constrain inputs to the marine-debris domain and filter implausible outputs.
Privacy. In the OceanGuard app, photos are analysed on-device and never leave the phone; the adapter itself processes no user data. Outside the app, operators are responsible for compliance with applicable privacy regulation.

How to Use

A. PEFT + Unsloth

python

from unsloth import FastVisionModel model, tokenizer = FastVisionModel.from_pretrained(    "unsloth/gemma-4-E2B-it",    load_in_4bit=True,)model.load_adapter(    "asferrer/gemma-4-E2B-it-oceanguard-marine-debris",    adapter_name="oceanguard",)model.set_adapter("oceanguard")FastVisionModel.for_inference(model)

from unsloth import FastVisionModel model, tokenizer = FastVisionModel.from_pretrained(    "unsloth/gemma-4-E2B-it",    load_in_4bit=True,)model.load_adapter(    "asferrer/gemma-4-E2B-it-oceanguard-marine-debris",    adapter_name="oceanguard",)model.set_adapter("oceanguard")FastVisionModel.for_inference(model)

B. llama.cpp (GGUF Q4_K_M)

See GGUF download. The OceanGuard Android app downloads and loads this file through its "Gemma 4 FT (GGUF)" provider using LlamaTextEngine (llama.cpp).

License and Citation

LoRA adapter: Apache 2.0.
Base model: google/gemma-4-E2B-it, subject to the Gemma Terms of Use. The merged GGUF is a derivative of the base model and is also subject to those terms.
Training data: source-dataset licenses apply (see the table in Training Data).

If you use this adapter, please cite the adapter, the application and the article:

bibtex

@misc{sanchezferrer2026oceanguard_adapter,  author       = {S{\'a}nchez-Ferrer, Alejandro},  title        = {{Gemma 4 E2B OceanGuard Marine Debris LoRA Adapter}},  year         = {2026},  howpublished = {Hugging Face Model Hub},  note         = {LoRA adapter for google/gemma-4-E2B-it. Apache 2.0.},  url          = {https://huggingface.co/asferrer/gemma-4-E2B-it-oceanguard-marine-debris}} @misc{sanchezferrer2026oceanguard_app,  author       = {S{\'a}nchez-Ferrer, Alejandro},  title        = {{OceanGuard AI: Fully Offline Marine Debris Intelligence with Gemma 4 and Two-Phase Tool Calling}},  year         = {2026},  howpublished = {Kaggle Gemma 4 Good Hackathon (Global Resilience track)},  url          = {https://github.com/asferrer/OceanguardAI-App}} @misc{sanchezferrer2026synthetic,  author       = {S{\'a}nchez-Ferrer, Alejandro and Gallego, Antonio Javier and Calvo-Zaragoza, Jorge},  title        = {{When Does Synthetic Data Help Underwater Debris Detection? Gap Metrics, Data Regimes and Detector Paradigms}},  year         = {2026},  howpublished = {Neurocomputing},  note         = {Manuscript in preparation}}

@misc{sanchezferrer2026oceanguard_adapter,  author       = {S{\'a}nchez-Ferrer, Alejandro},  title        = {{Gemma 4 E2B OceanGuard Marine Debris LoRA Adapter}},  year         = {2026},  howpublished = {Hugging Face Model Hub},  note         = {LoRA adapter for google/gemma-4-E2B-it. Apache 2.0.},  url          = {https://huggingface.co/asferrer/gemma-4-E2B-it-oceanguard-marine-debris}} @misc{sanchezferrer2026oceanguard_app,  author       = {S{\'a}nchez-Ferrer, Alejandro},  title        = {{OceanGuard AI: Fully Offline Marine Debris Intelligence with Gemma 4 and Two-Phase Tool Calling}},  year         = {2026},  howpublished = {Kaggle Gemma 4 Good Hackathon (Global Resilience track)},  url          = {https://github.com/asferrer/OceanguardAI-App}} @misc{sanchezferrer2026synthetic,  author       = {S{\'a}nchez-Ferrer, Alejandro and Gallego, Antonio Javier and Calvo-Zaragoza, Jorge},  title        = {{When Does Synthetic Data Help Underwater Debris Detection? Gap Metrics, Data Regimes and Detector Paradigms}},  year         = {2026},  howpublished = {Neurocomputing},  note         = {Manuscript in preparation}}

Sánchez-Ferrer, A., Valero-Mas, J. J., Gallego, A. J., & Calvo-Zaragoza, J. (2023). An experimental study on marine debris location and recognition using object detection. Pattern Recognition Letters, 168, 154-161. doi.org/10.1016/j.patrec.2022.12.019
Sánchez-Ferrer, A., Gallego, A. J., Valero-Mas, J. J., & Calvo-Zaragoza, J. (2022). The CleanSea Set: A Benchmark Corpus for Underwater Debris Detection and Recognition. IbPRIA 2022, Lecture Notes in Computer Science, Springer. doi.org/10.1007/978-3-031-04881-4_49
Sánchez-Ferrer, A. (2024). Modelos de difusión aplicados a la detección de objetos en el fondo marino. Master's Thesis, Universidad de Alicante. rua.ua.es
Sánchez-Ferrer, A. (2021). Deep Learning aplicado a la detección de residuos en el fondo marino. Bachelor's Thesis, Universidad de Alicante. rua.ua.es

Acknowledgments

Google DeepMind for releasing Gemma 4 under terms that enable open research and on-device deployment.
The Unsloth team for the FastVisionModel implementation that makes a Gemma 4 vision-language fine-tune tractable on a single workstation GPU.
CleanSea, Ocean_garbage and Neural_Ocean dataset authors for releasing marine-debris annotations.
Pattern Recognition and Artificial Intelligence (PRAI) group, Universidad de Alicante, which hosts the doctoral programme under which this work is carried out and co-authored the cited publications.

gemma-4-E2B-it-oceanguard-marine-debris

README

Files in This Repository

Model Description

Intended Use

Training Data

Training Procedure

Evaluation

Protocol

Configurations compared

Aggregate

Per-class mAP@0.5, TEST-200

Per-source mAP@0.5, TEST-200

Prompt used for detection

Deployment

GGUF download

Limitations

Bias, Risks and Safety

How to Use

A. PEFT + Unsloth

B. llama.cpp (GGUF Q4_K_M)

License and Citation

Related Publications by the Author

Acknowledgments

Explore FriendliAI today

README

Files in This Repository

Model Description

Intended Use

Training Data

Training Procedure

Evaluation

Protocol

Configurations compared

Aggregate

Per-class mAP@0.5, TEST-200

Per-source mAP@0.5, TEST-200

Prompt used for detection

Deployment

GGUF download

Limitations

Bias, Risks and Safety

How to Use

A. PEFT + Unsloth

B. llama.cpp (GGUF Q4_K_M)

License and Citation

Related Publications by the Author

Acknowledgments