Hugging Face model repository · Local preview
TagaLLM
Everyday Metro Manila Tagalog and Taglish chat.
Text generation Tagalog · tl Google Gemma 4 E2BModel details
The evaluated local app used original Google Gemma 4 E2B in Q4 with TagaLLM’s conversation prompt, cultural references, and language and scope checks. It is designed for everyday conversation. Long exchanges can lose track of updated details, and the language check can occasionally misclassify short replies.
The saved adapter training assets pin Google Gemma 4 E2B IT revision 4adf15748de9e1a0097b6274cee9c15836c41c37. The downloadable Q4 file matches the evaluated checkpoint checksum. Google lists Apache 2.0 for upstream Gemma 4. Google’s model card
Fine-tuning experiments
We trained experimental LoRA adapters on selected reviewed Tagalog dialogues. In the 22-dialogue TV pilot, held-out assistant loss fell from 4.44 to 3.72. Base Gemma and the adapted model each fully answered 2 of 6 final requests. The app configuration with cultural references answered 4 of 6 on that same development suite. The evaluated app and downloadable setup use base weights. The adapter is a separate research artifact.
Evaluation
Our development harness sends the next user message after each generated reply and records raw text, latency, relevance, naturalness, and state tracking. The saved E2B reference-enabled run passed all six exchanges in 3 of 20 chats under editorial scoring. See the article for denominators and examples.
A six-exchange pass requires all six replies to earn 2 out of 2 for naturalness and relevance, plus 2 out of 2 on every applicable state-tracking check. Reviewers record a note when state tracking does not apply. The harness requires a nonempty reply on every turn.
Across the 120 logged replies in the reference-enabled E2B development run, the local app took a median of 2.16 seconds per reply. The 95th percentile was 3.03 seconds. These timings come from the guarded development endpoint on the test machine.
Files and versions
| Artifact | Contents |
|---|---|
| Evaluated chat configuration | Local Gemma runtime, prompt, cultural references and language and scope checks. |
| Gemma culture adapter | Experimental LoRA pilot documented in the evaluation evidence. |
| Scratch research models | 3.75M / 256, approximately 15.5M / 256 and 15.52M / 512. The Hugging Face model cards provide the weights, tokenizers, manifests, and local runners. |
| Custom tokenizer | Download the original 2K byte-level BPE from the 256-token runs or the 4K BPE used in the later 512-token copy comparison. Gemma uses its own tokenizer. |
| Data and credits | The evidence page documents historical inputs and attribution. The Hugging Face dataset card describes the separate redacted candidate. |
Historical training datasets
The frozen 256-token snapshot contains 6,142 training conversations and 815 validation conversations. The 15.5M prose-first run also used 84 prose passages. The later 512-token copy comparison used a separate pinned mixture, documented with its starting checkpoint and input hashes. See the training data and checkpoint inventory
Try it locally
Follow the installation guide to run the chat UI with the downloadable Gemma model. The evaluated configuration also used SEA-LION for language and scope checks. The published local setup uses Gemma alone.
Browse model downloads · Gemma model card and files on Hugging Face · Find the saved runs behind these results