timothydillan/gemma4-e2b-balinese-assistant-v7

Balinese Gemma 4 E2B Assistant v7 LoRA

Experimental Balinese assistant LoRA adapter for timothydillan/gemma4-e2b-balinese-cpt.

This is a behavior-repair checkpoint after v6. It targets the remaining v6 smoke eval failures: assistant identity/greeting behavior, Bali/Denpasar factuality, and finite beginner dialogues.

Status

Research-only experimental checkpoint. Do not present this as a native-reviewed or production-ready Balinese assistant.

The v7 repair rows are AI-curated and pending native-speaker review.

Training

• Base: timothydillan/gemma4-e2b-balinese-cpt
• Dataset: data/processed/llm/sft_train_assistant_v7.jsonl
• Examples: 2,495
• v7 behavior-repair curated rows: 321
• v6 direct curated rows retained: 343
• MURI rows: 180
• Max sequence length: 1024
• Epochs: 2
• Learning rate: 3e-5
• LoRA rank/alpha: 16/16
• Final train loss: 0.4504
• Kaggle kernel: timothydillan/oim-balinese-gemma4-e2b-v0, version 14

Training completed without non-finite loss, traceback, or CUDA OOM in the captured Kaggle log.

Intended Use

Use with the base model through PEFT adapter loading for private research, qualitative evaluation, and low-resource Balinese assistant iteration.

Primary intended prompts include:

• Balinese and Indonesian questions about Balinese language/culture;
• concise greetings and small talk;
• beginner phrase help;
• short two-person example dialogues;
• direct explanations of common Bali terms such as canang sari, Nyepi, Galungan, Tri Hita Karana, Subak, and desa kala patra.

Limitations

• Not native-speaker reviewed.
• May still hallucinate, mistranslate, or mix registers.
• Cultural and religious answers need review by knowledgeable Balinese speakers.
• Not yet evaluated beyond a small smoke suite.
• Not intended for legal, medical, financial, religious authority, or public production use.

Next Evaluation Gate

This adapter should be judged against the hardened smoke eval in scripts/llm_assistant_smoke_eval.py.

Required improvements over v6:

• identity prompt answers as a Balinese assistant and avoids body/food/age drift;
• Bali prompt avoids the previous geography/capital confusion;
• food prompt stays loop-free;
• dialogue prompt avoids wrong English glosses and unfinished endings.

Reproducibility

Relevant local repo artifacts:

• src/oim/llm/assistant_v7.py
• scripts/build_assistant_sft_v7.py
• scripts/launch_kaggle_llm_assistant_v7.sh
• docs/data_audits/assistant_v7_data_decision.md