efficiencyx/Jun-LoRA-E2B-GGUF

Jun-E2B-GGUF

Merged GGUF builds of the Jun LoRA on Gemma 4 E2B (QAT) — a fine-tune trained on a compact, heavily curated synthetic conversational dataset derived from the visual novel My Dystopian Robot Girlfriend. The model captures the personality, speech patterns, and emotional nuance of the character Jun while preserving the base model's general reasoning and instruction-following capabilities.

This is the small sibling of Jun-12B, trained on the same dataset and the same output contract, sized to run on modest hardware.

The adapter is merged into the base weights here — these are standalone models, no --lora flag needed. (For E2B in particular, runtime --lora is not an option at all; see Notes.)

Model Variants & Repositories

Repository	Format	Description
efficiencyx/Jun-LoRA-E2B-GGUF	GGUF (Q8_0 / Q6_K / Q4_K_M)	Merged, quantized, for local inference
efficiencyx/Jun-LoRA-E2B-Adapter	LoRA Adapter	The adapter merged into these builds, currently private
efficiencyx/Jun-LoRA-12B-GGUF	GGUF	Larger sibling, same dataset
efficiencyx/Jun-v4-E2B-GGUF	GGUF	Previous generation (v4)

Quantization Guide

Quant	Size	Use Case
Q8_0	4.9 GB	Best quality, comfortable on 8 GB VRAM
Q6_K	3.8 GB	High quality, minimal loss
Q4_K_M	3.4 GB	Smallest footprint, fits 6 GB VRAM

Sizes are measured, not estimated. The base model is QAT (quantization-aware trained), so lower quants hold up better than a standard FP16 export. All three are quantized from the same bf16 master — no requantization chain, no imatrix.

The spread between Q6_K and Q4_K_M is small at this size, so there is little reason to drop below Q6_K unless VRAM is genuinely tight.

Jun-LoRA-E2B.BF16-mmproj.gguf (1.0 GB) is the multimodal projector — it carries both the vision (gemma4v) and audio (gemma4a) encoders from the base model. Download it only if you want image or audio input; text-only use does not need it. Pair it with any of the quants above.

These are unmodified Gemma 4 E2B weights: the LoRA targets the language tower only, so the perception encoders are stock. They ship at BF16 rather than quantized.

Usage

llama-server -m Jun-LoRA-E2B.Q6_K.gguf --jinja -ngl 99 -c 8192

--jinja is required. Without it llama.cpp ignores the embedded chat template and tool calls come back as plain text instead of structured calls.

With vision:

llama-mtmd-cli -hf efficiencyx/Jun-LoRA-E2B-GGUF --jinja

Intended Use

Conversational backend for Jun OS, an AI companion webapp:

• Character-consistent multi-turn conversation
• AI companion / interactive fiction applications
• Research into character-faithful fine-tuning on small, high-quality datasets

E2B exists for the case where 12B does not fit — local, low-VRAM, or latency-sensitive deployments.

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 itself (action tags, mood tags, tool calls) holds up well.

Training Details

Parameter	Value
Base model	unsloth/gemma-4-E2B-it-qat-q4_0-unquantized
Method	LoRA (rsLoRA)
LoRA rank	32
LoRA alpha	32
LoRA dropout	0.0
Target modules	q/k/v/o + gate/up/down (language tower)
Checkpoint	step 60
Framework	Unsloth

Why step 60 and not 3 epochs

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

Runtime LoRA is not possible on E2B. Gemma 4 E2B shares KV projections across layers 15–34, so a GGUF conversion of the base contains attn_k/attn_v only for layers 0–14. A LoRA trained on all layers cannot be applied with llama.cpp --lora, which fails on the first missing tensor. This is why these builds ship pre-merged.

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.