.litertlm build of the Jun LoRA on Gemma 4 E2B (QAT), for LiteRT-LM — Google AI Edge's on-device runtime (Android, desktop, embedded). Same fine-tune as the GGUF builds: a compact, heavily curated synthetic conversational dataset derived from the visual novel My Dystopian Robot Girlfriend, capturing the personality, speech patterns, and emotional nuance of the character Jun while preserving the base model's general reasoning and instruction-following.
The adapter is merged into the base weights — this is a standalone model. LiteRT-LM has no runtime adapter path for E2B anyway (see Notes).
int4 channelwise (embeddings + FFN/attention), int8 for the per-layer embedding projections
Activations
fp32
Embedder
externalized
Backend
CPU (XNNPack)
Exported with
litert-torch 0.9.3
This matches Google's own CPU distribution of E2B in size and recipe. Their 2.0 GB GPU/web variants use a 2/4/8-bit mixture that no published recipe reproduces; this build is 4/8-bit only.
Text-only. The LiteRT export covers the language tower — no vision or audio encoder, unlike the GGUF mmproj.
Usage
litert-lm run model.litertlm --prompt "Ciao Jun"
Or from the LiteRT-LM C++/Android APIs, pointing at the same file.
Supply a system prompt. Without one the model answers as stock Gemma — the persona lives behind the Jun OS system instruction, not in the weights alone. With the CLI:
litert-lm run model.litertlm --preset preset.py --prompt "Chi sei?"
The symbol must be lowercase system_instruction; SYSTEM_INSTRUCTION is silently ignored.
The bundled chat template is the stock Google Gemma 4 one, not Unsloth's — LiteRT-LM renders templates with minijinja, which has no map.get(), and Unsloth's template calls it 23 times.
Intended Use
Conversational backend for Jun OS, an AI companion webapp — specifically its on-device path:
Mobile / embedded deployment where a GGUF runtime is not an option
Research into character-faithful fine-tuning on small, high-quality datasets
Limitations
Specialized for a single character persona; not a general-purpose assistant.
Outputs reflect fictional narrative tropes and are not factual information or advice.
Performance degrades far outside the training distribution.
Inherits any biases present in the Gemma 4 E2B base weights.
At E2B scale the model is noticeably less coherent than Jun-12B: it can contradict itself inside a single reply and tends to lose the reply-length rule when asked for detailed explanations. The output contract (action tags, mood tags, tool calls) holds up well, including through int4 quantization.
Training was planned for 3 epochs but the released checkpoint is step 60. Evaluation loss bottomed out around 1.03 at step 60 and degraded afterwards — roughly 1.38 at step 70, recovering only partially to 1.16 at step 80. Later checkpoints did not recover the step-60 quality.
Behavioural probing agreed with the loss curve. The step-70 checkpoint in particular stopped responding to the live gauge values on physical-contact turns, emitting an identical mood update whether affection was 8 or 92 and whether tension was 10 or 90 — a collapse the step-60 checkpoint does not show. Step 60 was released on that basis.
Notes and Known Behaviour
The merge drops 40 of 245 LoRA pairs, by necessity. Gemma 4 E2B shares KV projections across layers 15–34, so the checkpoint contains k_proj/v_proj only for layers 0–14. Training wrapped the unused modules anyway; those deltas were never reachable at inference and are dropped at merge time. Every other pair is applied. This is the same structural fact that makes runtime llama.cpp --lora impossible on E2B, and why both this and the GGUF builds ship pre-merged.
Export needed patches. E2B's head_dim alternates 256/512 per layer, and litert-torch 0.9.3 infers a single global value — every KV cache came out 256 and attention failed. The export used a patched cache-shape inference reading per_layer_config[i].
Dataset
Synthetic conversational data derived from the visual novel My Dystopian Robot Girlfriend, curated for character consistency and for a structured output contract: inline [A:...] action tags, a trailing [A:mood_shift|...] bookkeeping tag, and tool calls. Roughly half of the assistant turns carry an explicit reasoning trace.
License
Apache 2.0, inherited from the base model. The character and source material belong to their respective owners; this fine-tune is a non-commercial fan project.
