lbartoszcze/jeden-goal-qwen3-4b

Jeden Goal Qwen3 4B

Jeden Goal Qwen3 4B is a task-specific Qwen3-4B model fine-tuned to turn a coding-agent request into a concise 3–7 word imperative goal. It preserves product names and technical identifiers, follows the user's language, and emits only <goal>…</goal> or <goal/> when a continuation has no self-contained task.

The repository contains the Q4KM GGUF used by Jeden Desktop. Inference is local through llama.cpp; transcript text is not sent to Hugging Face or another inference service. Jeden Desktop downloads the immutable model once, verifies its SHA-256, and caches both the model and generated goals on the Mac.

Artifact

Prompt contract

Use the system prompt in `goal-system-prompt.md`, then provide the request as:

<user>the login button is broken on mobile somehow, can you fix?</user>

Expected output shape:

<goal>Fix login button on mobile</goal>

A context-dependent continuation has no self-contained task:

<user>yes, do that</user>
<goal/>

Run locally with a recent llama-cli:

llama-cli \
  --model jeden-goal-qwen3-4b-q4_k_m.gguf \
  --system-prompt-file goal-system-prompt.md \
  --prompt '<user>sprawdź aplikacje desktopowe Brama i Skarbiec</user>' \
  --single-turn --reasoning off --ctx-size 2048 \
  --n-predict 40 --temp 0 --no-display-prompt --simple-io

Training

The model was fine-tuned for 3 epochs at a learning rate of 1e-5 on 564 curated multilingual goal pairs derived from privacy-masked coding-agent sessions. The raw transcripts and training rows are private and are not included in this repository.

The held-out split contains 51 manually curated rows. Training evaluation recorded:

• —exact match: 2/51 (3.92%);
• —final evaluation loss: 0.5859;
• —independent semantic audit: 51/51 outputs judged sensible;
• —nonsensical or unparseable outputs: 0.

Exact match is intentionally strict: semantically equivalent language changes such as Dodaj CLI i MCP do Weles versus Add CLI and MCP to Weles count as failures. The semantic audit is therefore the release qualification criterion, while exact match remains visible as a diagnostic.

Limitations

• —The model formulates goals; it is not a general assistant and should not be used to answer the request.
• —Polish inputs can occasionally produce English goals.
• —It can slightly broaden scope, for example changing “identify the crash cause” into “identify and resolve the crash.”
• —It only sees the supplied request. Context-dependent continuations should produce <goal/> rather than infer missing context.
• —The model can reproduce biases or mistakes from Qwen3-4B and the private distillation labels.

Provenance and license

The qualified GGUF is content-addressed by the SHA-256 above. Training and qualification are owned by `wisent-ai/transcript-label-trainer`; runtime integration is owned by `wisent-ai/jeden-desktop`.

This derivative follows the Apache 2.0 license of Qwen/Qwen3-4B. See `LICENSE`.