CoolFace
Modelpublic

xxue752/mental_health_best_model_litert

sourceHugging Facegemmaupdated 5mo agoView on Hugging Face
1likes10downloads
Model Card

mentalhealthbestmodellitert

LiteRT (TFLite) INT4-quantized export of the multitask mental-health LoRA-finetuned Gemma-4-E4B-it. Designed for on-device / edge deployment via LiteRT-LM and TFLite runtimes.

Disclaimer — research use only. The outputs are predicted class IDs from a frozen 1-of-K head built into a chat-style prompt; they are not clinical advice.

Source / lineage

StageRepo
Base`google/gemma-4-E4B-it`
LoRA adapter (140 MB)`xxue752/mental_health_best_model`
Merged bf16 (~16 GB)`xxue752/mental_health_best_model_merged`
LiteRT INT4 (this repo)xxue752/mental_health_best_model_litert

Files

FileSizePurpose
model.litertlm4.12 GBSingle-file LiteRT-LM bundle (preferred) — load with litert-lm runtime
model_quantized.tflite2.2 GBPrefill + decode TFLite (dynamic_wi4_afp32)
embedder_quantized.tflite327 MBExternal token embedder (Gemma 4 requires externalize_embedder=True)
per_layer_embedder_quantized.tflite1.4 GBPer-Layer Embeddings (Gemma 4 PLE)
tokenizer.json, tokenizer_config.json32 MBGemma 4 tokenizer
chat_template.jinja17 KBChat template (LiteRT-LM-compatible variant — see "Compatibility fix" below)
llm_metadata.pb17 KBLiteRT-LM runtime metadata (carries the patched chat template)