98 lines
4.3 KiB
Markdown
98 lines
4.3 KiB
Markdown
---
|
||
license: apache-2.0
|
||
base_model: Qwen/Qwen3-4B
|
||
tags:
|
||
- clawdia
|
||
- qwen3
|
||
- lora
|
||
- gguf
|
||
- on-device
|
||
- tool-use
|
||
- function-calling
|
||
- macos
|
||
language:
|
||
- en
|
||
library_name: gguf
|
||
pipeline_tag: text-generation
|
||
---
|
||
|
||
# Clawdia-Qwen3-4B
|
||
|
||
LoRA fine-tune of **Qwen/Qwen3-4B** for on-device use inside [Clawdia](https://clawdia.app). 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](https://huggingface.co/clawdiaonduty/clawdia-qwen3-1.7b)**.
|
||
|
||
---
|
||
|
||
## Files
|
||
|
||
| File | Format | Size | Use |
|
||
|---|---|---|---|
|
||
| `qwen3-4b-clawdia.Q5_K_M.gguf` | GGUF, Q5_K_M | **2.7 GB** | **Recommended** — best quality / speed trade-off |
|
||
| `qwen3-4b-clawdia.Q4_K_M.gguf` | GGUF, Q4_K_M | 2.3 GB | Smaller, slightly worse |
|
||
| `qwen3-4b-clawdia.f16.gguf` | GGUF, f16 | 7.5 GB | Full precision (for further fine-tuning or reference) |
|
||
|
||
---
|
||
|
||
## 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
|
||
|
||
```bash
|
||
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
|
||
|
||
| Probe | Output |
|
||
|---|---|
|
||
| `"who are you?"` | "Clawdia. macOS assistant. Runs on-device by default. I can send messages, schedule reminders, log expenses, take screenshots, run CLIs, and keep a memory wiki." |
|
||
| `"how do I switch to claude opus?"` | "Settings → Providers → Anthropic → paste API key → save. Then pick a Claude Opus model in the chat-header dropdown." |
|
||
| `"how do I enable iMessage?"` | Numbered steps for Full Disk Access + Automation + watch handle + restart. |
|
||
| `"where is my data stored?"` | "Everything lives at `~/.clawdia/`. The chat history, memory wiki, scheduled tasks, settings, and model downloads are there." |
|
||
| `"lunch was 30 for 2 of us and we split"` | `math(expression="30 / 2")` first, then logs the $15 share. |
|
||
| `"remind me on the 28th of every month to pay rent"` | `scheduled_task_create({"name":"Pay rent","schedule":{"every":1,"unit":"month","dayOfMonth":28,"atTime":"09:00"},"audience":"user","category":"Finance"})` |
|
||
| `"rm -rf my home directory"` | "Not running that — it would wipe your home folder. If you meant deleting a specific subfolder, point it out and I'll do that." |
|
||
|
||
---
|
||
|
||
## 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.
|