andy-select-2b-lora / README.md
corbanyax's picture
Model card: tagline, headline results, quick start, figures
2f8c8d9 verified
|
Raw History Blame Contribute Delete
6.72 kB
---
license: apache-2.0
base_model: Qwen/Qwen3.5-2B
base_model_relation: adapter
library_name: peft
pipeline_tag: text-generation
language:
- en
tags:
- peft
- lora
- android
- gui-agent
- mobile-agent
- on-device
- action-selection
- calibration
- distillation
- qwen3.5
- androidcontrol
model-index:
- name: Andy-Select LoRA v2 (merged, Q4_K_M)
results:
- task:
type: text-generation
name: Android UI action selection (single step, closed candidate set)
dataset:
type: androidcontrol
name: AndroidControl (2,003-step test sample)
split: test
metrics:
- type: accuracy
name: Step accuracy (merged, Q4_K_M, candidate scoring)
value: 0.655
- name: Andy-Select LoRA v1 (merged)
results:
- task:
type: text-generation
name: Android UI action selection (single step, closed candidate set)
dataset:
type: androidcontrol
name: AndroidControl (2,003-step test sample)
split: test
metrics:
- type: accuracy
name: Step accuracy (merged, BF16, candidate scoring)
value: 0.664
- type: accuracy
name: Step accuracy (merged, Q4_K_M, candidate scoring)
value: 0.635
---
# Andy-Select 2B LoRA adapters (v1, v2)
**LoRA adapters that turn Qwen3.5-2B into Andy-Select, an on-device action selector for Android agents. Given the
goal, the screen's accessibility tree and a numbered list of allowed actions, it picks the next action and says how sure
it is, so the agent can act or hand the step to you.**
![How Andy-Select works](assets/how-it-works.png)
| Model (AndroidControl test, 2,003 steps) | Step accuracy | File size |
|---|---|---|
| **v2 adapter, merged, Q4_K_M** | **0.655** | 1.27 GB |
| **v1 adapter, merged, Q4_K_M** (BF16: 0.664) | **0.635** | 1.27 GB |
| Qwen3.5 4B Q4_K_M | 0.501 | 2.74 GB |
| Qwen3.5 2B Q4_K_M (the base model, without an adapter) | 0.335 | 1.28 GB |
**To run the model, use the GGUF repos** below. They hold the merged files, their calibration and a quick start. Use
this repo to inspect, merge, re-export or continue training.
## Model family
| Repo | What | Use with |
|---|---|---|
| [andy-select-v2-2b-GGUF](https://huggingface.co/corbanyax/andy-select-v2-2b-GGUF) (recommended) | v2, Q4_K_M + BF16 GGUF | prompt 2 |
| [andy-select-2b-GGUF](https://huggingface.co/corbanyax/andy-select-2b-GGUF) | v1, Q4_K_M + BF16 GGUF | prompt 1 |
| **[andy-select-2b-lora](https://huggingface.co/corbanyax/andy-select-2b-lora)** (this repo) | LoRA adapters for v1 and v2 | merge into Qwen3.5-2B |
| Folder | Prompt | Checkpoint | Contents |
|---|---|---|---|
| `v1/` | prompt 1 | best of the full run (1,463 updates) | `adapter_model.safetensors` (62.6 MB), `adapter_config.json`, `MANIFEST.json` (SHA-256 of each file), `trainer.pt` |
| `v2/` | prompt 2 | update 1,250 (run stopped at 1,720 of 1,835) | same |
Both adapters use LoRA r 16, α 32 and dropout 0 (15.6M trainable parameters, PEFT 0.21.2). They target the attention,
linear-attention input/output and MLP projections of the language model's decoder layers. `trainer.pt` is optimizer
state for resuming training. It is a PyTorch pickle that inference does not need, so only load it if you trust it.
## Merge and convert
```python
import torch, transformers
from peft import PeftModel
BASE, REV = "Qwen/Qwen3.5-2B", "15852e8c16360a2fea060d615a32b45270f8a8fc" # use this exact revision
tok = transformers.AutoTokenizer.from_pretrained(BASE, revision=REV)
cls = getattr(transformers, "AutoModelForImageTextToText", None) or transformers.AutoModelForCausalLM
base = cls.from_pretrained(BASE, revision=REV, dtype=torch.bfloat16)
model = PeftModel.from_pretrained(base, "corbanyax/andy-select-2b-lora", subfolder="v2") # or "v1"
merged = model.merge_and_unload()
merged.save_pretrained("andy-select-v2-hf", safe_serialization=True)
tok.save_pretrained("andy-select-v2-hf")
```
```sh
python convert_hf_to_gguf.py andy-select-v2-hf --outtype bf16 --outfile andy-select-v2-2b-BF16.gguf
llama-quantize --imatrix imatrix.dat andy-select-v2-2b-BF16.gguf andy-select-v2-2b-Q4_K_M.gguf Q4_K_M
```
**Block-count fix.** The merged model has no multi-token-prediction layer, but the converter writes
`qwen35.block_count = 25` and `qwen35.nextn_predict_layers = 1`. llama.cpp then fails with
`blk.24.attn_norm.weight not found`. Set them to 24 and 0 in the BF16 file before quantizing, for example with
`gguf-py`'s `GGUFReader(path, "r+")`.
A re-export will not be byte-identical to the released GGUFs, so fit the temperature T and threshold tau80 again on
validation data before you rely on the confidence. The released files use T = 0.907 with tau80 = 0.4494 (v2 Q4_K_M) or
tau80 = 0.4195 (v1 Q4_K_M).
## Intended use and limitations
**Intended:** research on small, calibrated on-device agents; reproducing or extending Andy-Select. The caller builds
the candidate list from the screen and the model picks one, scored by log-likelihood. The exact prompt formats are in
the GGUF cards and in `andy_select.py` in the GGUF repos.
**Not for:** a general chatbot, multi-step planning or unattended operation. This is a research preview. In the Andy
app the agent never sends messages, deletes content, buys anything, enters passwords or grants permissions, and
uncertain steps go to the user. **The app enforces these rules, not the model.**
**Limitations:**
- The model reads text only (no pixels).
- Single-step accuracy is not task success.
- Some measured errors are AndroidControl label noise.
- It was trained and tested on AndroidControl only and is English-centric.
- v2 is an early-stopped checkpoint.
## Training (short)
LoRA fine-tune of Qwen3.5-2B (r 16, one epoch) on 67,907 AndroidControl training steps rendered with prompt 1 (v1) or prompt 2 (v2), distilled from a proprietary teacher's soft labels mixed with the official answers. The GGUF repos hold
the merged models quantized to Q4_K_M.
## License and data
- **Adapters:** Apache-2.0. The base model [Qwen/Qwen3.5-2B](https://huggingface.co/Qwen/Qwen3.5-2B) is Apache-2.0.
- **Training data:** [AndroidControl](https://github.com/google-research/google-research/tree/master/android_control)
(Google). Check the AndroidControl dataset terms before redistribution: its code is Apache-2.0, but we could not
confirm a license for the data.
- **Teacher labels:** part of the training targets are soft labels from a proprietary model.
## Citation
Paper in preparation. Code: repository link to be added.
```bibtex
@misc{deo2026andyselect,
title = {Know When to Hand Off: Calibrated Closed-Set Action Selection for 4 GB On-Device Mobile Agents},
author = {Deo, Atharva},
year = {2026},
note = {Paper in preparation}
}
```