clawdia-qwen3-4b

Clawdia-Qwen3-4B

LoRA fine-tune of Qwen/Qwen3-4B for on-device use inside Clawdia. Bigger sibling of the 1.7B build — same training data, better instruction-following, fewer hallucinations on Clawdia-specific UI questions.

This is the recommended local model for systems with 8+ GB RAM. The Q5_K_M GGUF is ~2.7 GB; pair it with Clawdia's bundled llama.cpp runtime.

For a smaller (1.2 GB) variant, see clawdiaonduty/clawdia-qwen3-1.7b.

---

Files

---

How to use

Inside Clawdia (recommended)

Settings → Local Inference → pick Clawdia-Qwen3 4B Q5_K_M. Clawdia downloads to ~/.clawdia/local-inference/models/ and runs it via the bundled llama.cpp runtime.

llama.cpp directly

llama-completion \
  --model qwen3-4b-clawdia.Q5_K_M.gguf \
  --jinja \
  -sysf system_prompt.txt \
  -p "log $14.50 for lunch /no_think" \
  --temp 0.0 -n 280

Two critical flags:

• --jinja: enables the embedded Qwen3 chat template (tool calls depend on this).
• Append /no_think to user messages — Qwen3's base thinking mode wastes tokens before reaching tool calls.

---

What it does well

---

Training

• Base: Qwen/Qwen3-4B
• Adapter: LoRA rank 32, alpha 32, dropout 0.05 — applied to q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj on the top 16 transformer layers
• Data: 1,662 hand-authored multi-turn dialogs across 25 categories (finance, memory, iMessage, Telegram/WhatsApp, scheduled tasks, pantry, proactive, todos/habits/journal, setup/safety, edge cases, indirect/proactive offers, goal-aware reasoning, math splits, packages/orders, web/news, MCP tools, memory CLI, Clawdia self-knowledge, Clawdia UI / don't-lie discipline)
• Mask: train_on_responses_only — loss only on assistant tokens
• Schedule: AdamW, lr 2e-4, cosine decay, 5% warmup, 4 epochs (~430 steps), effective batch 16, max_seq_length=6144
• Hardware: 1× Modal H100, ~29 min wall-clock
• Loss: averaged 0.40 (train), best eval 0.565 at epoch 1.92 (final eval climbed — slight overfit; use earlier checkpoint if needed)

---

Known rough edges

• Tool-name drift in some finance/memory calls: occasionally emits finance_add_expense instead of canonical finance(action="add_expense"). Less frequent than the 1.7B variant but still happens. Targeted fix in next iteration.
• Identity string drift: When asked "what model are you?" the 4B variant still answers "Clawdia-Qwen3-1.7B" — the training data was authored for the 1.7B build. Cosmetic.

---

License

Apache 2.0 — inherited from Qwen/Qwen3-4B.