You can deploy this model and start earning money today!

carryer123/
gemma4-trilingual-family-Q4_K_M

Name: gemma4-trilingual-family-Q4_K_M
Availability: InStock

Views

No views yet

gemma4-trilingual-family-Q4_K_M

GGUF (Q4_K_M, ~3.2 GB) of a Gemma 4 E2B base with a state-gated multilingual family-tutor LoRA fused in. Designed to ship inside an iPad SwiftUI app (llama.cpp via xcframework) for Korean / English / Russian / French multi-generational households.

Companion code repo: carryer123/gemma4-trilingual-family. Submitted to: Gemma 4 Good Hackathon (Kaggle 2026).

Intended use

Drop into the app's Documents/ directory; the app auto-loads the first .gguf it finds. Running it raw via llama.cpp works for benchmarking but the runtime gate suite that audits every generation is implemented in the SwiftUI app, not in the model file.

Training recipe

Base: Gemma 4 E2B (instruction-tuned), bf16
LoRA: rank 32 / alpha 64 on q/k/v/o + gate/up/down
Curriculum: bridge-pivot trilingual triples + policy/family slices (schema, script-state, age-policy, session-routing exemplars)
1500-step single-arm training, seed 10 winner of a 3-seed sweep
Audit: frozen 112-probe four-language gate suite (G1 / G2 / G3 / G4)

Full recipe + audit JSONs in the companion GitHub repo.

Evaluation (4-language, KO / RU / FR / EN)

Metric	Translation-only baseline (n=2)	This adapter (n=3)
Held-out loss	0.804	0.667
G2 — script state	92.5 %	91.7 %
G3 — JSON schema	0.0 %	95.8 %
G4 — session routing	0.0 %	91.7 %

Cross-base replication on Qwen 2.5 3B / Llama 3.2 3B / Phi-3.5 mini recovers G3 to ≥ 91.7 % on every base; Llama additionally recovers G4 to 100 %. FP16-inference parity on the seed-10 winner matches the bf16 numbers within rounding (G3 = 100, G4 = 100, G2 = 92.5).

Quick try via llama.cpp

bash

huggingface-cli download carryer123/gemma4-trilingual-family-Q4_K_M \
  gemma4_e2b_policy.Q4_K_M.gguf --local-dir .
./llama-cli -m gemma4_e2b_policy.Q4_K_M.gguf \
  -p "추석에 송편을 만드는 이유를 아이 눈높이로 설명해 주세요." -n 256

Licensing

This adapter inherits the Gemma 4 Terms of Use from the base model. The training code and audit protocol around it are released under Apache 2.0 / CC-BY 4.0 (see the companion GitHub repo).