anjadhe-qwen3.5-4b

Anjadhe 4B (Qwen3.5-4B fine-tune)

The default local model for the Anjadhe personal AI app on Macs with 8–16 GB of memory. A LoRA fine-tune of Qwen/Qwen3.5-4B (Apache 2.0), tuned for the specific jobs Anjadhe runs all day on a small machine: reading email into structured insights, tool-calling against the app's task/goal/schedule/notes APIs, and answering questions about the user's own data — grounded, dated, and honest about what it can't find.

What's in the training data — and what isn't

Every training sample is synthetic. Emails come from seeded template generators; assistant conversations come from a scenario harness driving the real app against seeded demo data, with a larger open-weight model (Qwen3.6-35B-A3B) as the teacher. Labels were rejection-sampled hard: schema violations, invented dates (any date not literally present in the source), inconsistent action items, and unverified outcomes were dropped, never repaired. No user data of any kind was used. Prompt-injection attempts (instructions embedded in email bodies) are deliberately included with correct refusals, because a model that reads email is a model that reads attacker text.

Evaluations

Scored at temperature 0 through llama.cpp on the quantized artifact — the exact form that ships. Baselines are the untuned base at the same quant.

Intended use and limitations

Built to be good at Anjadhe's jobs at 4B size and speed — not a general-purpose replacement for larger models. Long, complex structured generations and deep open-ended reasoning remain better served by larger local models or hosted options, which Anjadhe offers as explicit opt-ins. The tune is coupled to Anjadhe's production prompts; it should behave like a normal Qwen3.5-4B on generic chat, but its improvements concentrate on Anjadhe-shaped inputs.

Files

• anjadhe-qwen3.5-4b-*-Q4_K_M.gguf — the shipping quantization (~2.6 GB)

Training pipeline (generators, teacher labeling, verification harness, training scripts) is documented in the Anjadhe project repos.