clawdia-qwen3-1.7b

Clawdia-Qwen3-1.7B

LoRA fine-tune of Qwen/Qwen3-1.7B for on-device use inside Clawdia — Anthropic-style assistant for macOS with native tool calling, persistent memory, scheduled tasks, finance/iMessage/Telegram integrations, and a browser surface.

Goal: be a fast (~1.2 GB) local model that handles the 20% of agent traffic that covers 80% of Clawdia use cases — finance logging, memory ops, scheduled reminders, messaging routing, math-derived splits, proactive nudges from stored goals.

---

Files

---

How to use

Inside Clawdia (recommended)

Settings → Local Inference → pick Clawdia-Qwen3-1.7B. Clawdia downloads it to ~/.clawdia/local-inference/models/ and runs it via the bundled llama.cpp runtime.

llama.cpp directly

llama-completion \
  --model qwen3-1p7b-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: tells llama.cpp to apply the embedded Qwen3 chat template (the model's tool-call format expects this).
• Append /no_think to user messages (or set enable_thinking=false via chat-template kwargs). The base Qwen3 has a thinking-mode chain-of-thought that this fine-tune doesn't use — leaving it on makes the model ramble before reaching the tool call.

Python with llama-cpp-python

from llama_cpp import Llama
llm = Llama(model_path="qwen3-1p7b-clawdia.Q5_K_M.gguf", n_ctx=4096, chat_format="chatml")
out = llm.create_chat_completion(
    messages=[
        {"role": "system", "content": SYSTEM_PROMPT},
        {"role": "user", "content": "log $14.50 for lunch /no_think"},
    ],
    temperature=0.0,
)
print(out["choices"][0]["message"]["content"])

---

Tool-call format

Trained to emit one tool call per assistant turn, wrapped in <tool_call>...</tool_call>:

<tool_call>
{"name": "finance", "arguments": {"action": "add_expense", "amount": "$14.50", "description": "lunch", "category": "food", "date": "2026-05-18"}}
</tool_call>

The next user turn must contain the tool result wrapped in <tool_response>...</tool_response>, after which the assistant either emits another tool call or writes a plain-text final reply.

---

What it does well

---

Training

• Base: Qwen/Qwen3-1.7B
• 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: 1,203 hand-authored multi-turn dialogs across 19 categories (finance log/analysis, memory write/read, iMessage, Telegram/WhatsApp, scheduled tasks, pantry, proactive multi-tool, todos/habits/journal, setup/safety, edge cases, indirect/proactive offers, goal-aware reasoning, math-derived expenses)
• Mask: train_on_responses_only — loss only on assistant tokens
• Schedule: AdamW, lr 2e-4, cosine decay, 5% warmup, 4 epochs (~300 steps), effective batch 16, max_seq_length=4096
• Hardware: 1× Modal H100, ~9 min wall-clock
• Loss: 3.46 → 0.27 (train), best eval 0.55 at epoch 2.6
• Export: mlx_lm-free path: HF safetensors → llama.cpp convert_hf_to_gguf.py → llama-quantize for Q4/Q5

---

Known rough edges (v1)

These are being addressed in v2:

• Occasional tool-name drift (~15% of finance/memory calls): invents finance_add_expense instead of canonical finance(action="add_expense").
• Hallucinated MCP tool names: e.g. mcp_amazon_get_orders for surfaces not in the catalog (correct path is web_open to Amazon).
• Schedule arg shape sometimes flat ({runAt: ...}) instead of nested ({schedule: {unit: "once", runAt: ...}}).
• Inbound iMessage replies occasionally repeat words — needs more varied training data.

---

License

Apache 2.0 — inherited from Qwen/Qwen3-1.7B.

---

Acknowledgments

• Built on Qwen/Qwen3-1.7B by Alibaba.
• Trained via unsloth + TRL on Modal Labs.
• Quantized via llama.cpp.
• For Clawdia — a personal on-device assistant for macOS.