mts-the-architect/open-sensations-gemma4-e2b-bf16-lora

open-sensations: Gemma 4 E2B text classifier

Separate Gemma-based Open classifier with the same task labels and per-window JSON contract as its RoBERTa counterpart. Adapter + classifier head download, not standalone base weights. Run with gemma_classifier.py; a generic chat/generation pipeline does not implement this classifier.

Selected Gemma held-out test micro/macro F1: 0.7821 / 0.7556. RoBERTa reference on the same split: 0.5772 / 0.5530. These are classification F1 scores, not accuracy or a general intelligence rating. Different adaptation budgets and model sizes limit architecture-only conclusions.

Pinned base: google/gemma-4-E2B at d29ff6b45f081a49ee2733a859c9c9c2d95d1a6f. The loader reads only text-backbone tensors; no vision/audio encoders, generative head, tools or chat outputs. Actual parameters including adapters/head: 4,629,917,958; trainable: 1,348,614. E2B is Google's effective-size designation, not the full parameter count. BF16 backbone; FP32 head and trainable adapters. Default root is the synthetic arm. The human baseline and both continuations are also separate ZIP downloads.

Labels in exact order: vision, hearing, smell, taste, touch, interoception. Input format: Target: WORD Context: SENTENCE. For sensations, the target word must be identified: Target: WORD followed by newline and Context: SENTENCE. This remains a small target-word sensory association prototype, not a general sentence-level sensation detector.

Training, hardware and reproducibility

ASUS ROG Flow Z13 2025 GZ302EA; Ryzen AI Max+395; Radeon8060S; 128GB unified RAM; Ubuntu24.04 WSL2. One local GPU; no cloud training. Settings: {"seed": 20261004, "epochs": 10, "microbatch_size": 8, "gradient_accumulation_steps": 4, "effective_batch_size": 32, "max_length": 128, "lora_rank": 8, "lora_alpha": 16, "lora_dropout": 0.05, "lora_targets": ["q_proj", "v_proj"], "learning_rate": 0.0001, "head_learning_rate": 0.001, "continuation_epochs": 1, "continuation_learning_rate": 5e-05, "continuation_head_learning_rate": 0.0005, "precision": "BF16 frozen backbone; FP32 adapters/head", "optimizer": "AdamW", "weight_decay": 0.01, "warmup_ratio": 0.06, "max_grad_norm": 1.0, "attention": "eager", "pooling": "last_non_padding_token"}. Versions: {"torch": "2.13.0+rocm10.0.0", "transformers": "5.18.0", "peft": "0.21.2", "accelerate": "1.15.0", "numpy": "2.5.3"}.

Human and evaluation data preserve the original RoBERTa splits. Full metadata, source hashes and pinned dataset revisions are in data_manifest.json and run.json; per-epoch histories, synthetic provenance and replay IDs are preserved. Raw human texts are not redistributed. Synthetic data is assistant-authored and not independently human-reviewed. Both continuation arms replay the same sampled human data; synthetic adds examples/compute. See learning guide.

Evaluation

open-sensations: Gemma results

Root adapter selected using validation only, before test evaluation. F1 is not accuracy.

Arm	Validation micro F1	Validation macro F1	Test micro F1	Test macro F1	Threshold
baseline	0.7360	0.6793	0.8088	0.7197	0.50
real-only	0.7229	0.6766	0.7674	0.7257	0.40
synthetic	0.7317	0.6967	0.7821	0.7556	0.30

Selected: synthetic. Synthetic accepted: True.

RoBERTa reference on the same held-out data

Micro F1: 0.5772; macro F1: 0.5530. Different tokenizer/model size/precision/training budgets: practical comparison, not architecture-only ablation. These test results did not select an arm.

Selected label	Precision	Recall	F1	Support
vision	0.7917	1.0000	0.8837	38
hearing	0.3750	0.8571	0.5217	7
smell	1.0000	1.0000	1.0000	6
taste	1.0000	1.0000	1.0000	6
touch	0.5500	0.8462	0.6667	13
interoception	0.3333	0.7500	0.4615	4

Full experiment: 18.05 minutes. Peak PyTorch allocated GPU memory: 9.69 GiB; reserved: 9.90 GiB. These are not total process/system memory.

Arm	Epoch	Mean training loss	Validation micro F1	Validation macro F1
baseline	1	0.79388	0.4956	0.2977
baseline	2	0.58331	0.5686	0.4946
baseline	3	0.50667	0.6914	0.6260
baseline	4	0.49347	0.7360	0.6793
baseline	5	0.48771	0.6818	0.6417
baseline	6	0.48051	0.6966	0.6637
baseline	7	0.47659	0.7006	0.6685
baseline	8	0.47428	0.7812	0.6731
baseline	9	0.47279	0.7383	0.6537
baseline	10	0.47260	0.6854	0.6438
real-only	1	0.49174	0.7229	0.6766
synthetic	1	0.44660	0.7317	0.6967

Download and classify

Use the pinned model-library environment described in requirements-gemma-wsl.lock.txt, with hardware-appropriate PyTorch. Download this repository's adapter/head/tokenizer/config files with huggingface_hub.snapshot_download('mts-the-architect/open-sensations-gemma4-e2b-bf16-lora', local_dir='./downloaded-task'). The custom loader also obtains the pinned public base (approximately 10.25 GB once in cache). It instantiates only the text backbone; adapters are not autonomous models.

import sys
sys.path.insert(0, './downloaded-task')  # Find the downloaded custom loader.
from gemma_classifier import load_classifier
import torch
model, tokenizer, options = load_classifier('./downloaded-task', device='cuda')
model.eval()
text = "Target: aroma\nContext: The aroma of bread filled the bakery."
inputs = tokenizer(text, return_tensors='pt', truncation=True, max_length=options['max_length'])
with torch.inference_mode():
    scores = model(**inputs).logits.sigmoid()[0].cpu().tolist()
labels = [model.config.id2label[i] for i, s in enumerate(scores) if s >= options['global_threshold']]

For the existing window evidence format, use python ./downloaded-task/predict_windows.py --task open-sensations --model ./downloaded-task --file document.txt --target aroma. Windowing guide documents overlap and uncalibrated max/mean summaries. API instructions describe the same routes, owner approvals and rate limits; pass --model-dir open-sensations=./downloaded-task. Uploading these files does not deploy a public HTTP service. The private API database is never included.

Limits and attribution

Research models: restricted data coverage, label disagreement, rare-category instability and small sensory supports. Gemma size does not guarantee higher classification quality or explicit reasoning. Sigmoid outputs are not calibrated probabilities; short-input thresholds do not establish document-level decisions. No model judges moral truth or measures actual sensations. Training/validation selection is independent of test results.

Base weights: Google Gemma 4, Apache 2.0; official card. Human data: GoEmotions, Apache 2.0; MFRC, CC BY 4.0; Contextualized Sensorimotor Norms, CC BY 4.0. Dataset licenses are distinct from model/code licensing. See LICENSE/NOTICE and source references in the learning guide.