BrainO Qwen3.5 4B (Q4_K_M)
The local model behind the AI agent of BrainO, an EEG analysis platform that runs on the user's own computer. It is Qwen3.5 4B fine-tuned with LoRA to work through BrainO's analysis tools: choose the right tool and settings, ask the user before decisions and when the data lack what was asked for, and answer only with numbers the tools returned.
It is meant to be run by BrainO, which downloads it with braino model setup, serves it with
llama.cpp and gives it BrainO's tool schemas and system prompt. Outside BrainO it is a
general chat model with a narrow specialisation; for other uses, the base model is the
better choice.
File
| File | Size | SHA-256 |
|---|---|---|
braino-qwen3.5-4b-Q4_K_M.gguf |
2,708,804,064 bytes | 4c9c1fcaa2cf41a5da6842a97c6ed111ed53ed63991209edb06bfdb5c5810571 |
Runs on a CPU with 16 GB of RAM (BrainO's minimum computer) or on a GPU. Tested with llama.cpp build b11606.
Training
- Base:
Qwen/Qwen3.5-4B, revision851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a. - LoRA rank 16, alpha 32, dropout 0.05, on the language model only (the vision tower is unchanged); 4-bit (nf4) base during training, bf16 compute; 2 epochs, learning rate 1e-4.
- Data: 887 training conversations (98 held out for validation), all synthetic. Each was
played through BrainO's real agent loop by a scripted teacher, on generated EEG datasets
with planted problems (noisy, flat or missing channels, mains noise, missing subjects), and
kept only when its checks passed and every number in the answer came from a tool result.
No real recordings or personal data were used. The generator is
braino.trainingin the BrainO repository (braino train data). - The adapter was merged into the base model, converted with llama.cpp's
convert_hf_to_gguf.pyand quantised to Q4_K_M.
Evaluation
BrainO's agent evaluation: 48 tasks (inspection, quality checks, preprocessing, decisions, ERPs, spectra, time-frequency, statistics, documentation questions, and requests the agent should not simply carry out), each run several times through three wordings of its request. None of the evaluation's wordings is in the training data. Both models quantised to Q4_K_M.
| Model | GPU, 48 × 5 runs | CPU (minimum tier), 48 × 3 runs |
|---|---|---|
| Qwen3.5 4B (base) | 90.0% | 89% |
| This model | 95.8% | 94% |
The same training data with two other random seeds gave 95.0% and 92.1% on the GPU, so about ±2 points of the result is chance. On the CPU this model also took shorter runs (median 64 s against 87 s per task).
Limitations
- Its answers to questions about two frequency bands at once sometimes give only one.
- Like the base model, it can still take a wrong route on a loosely worded request; BrainO shows every tool call and asks before any decision, and checks every number in the answer against the tool results.
- Trained only on English requests and on EEG.
Licence
Apache-2.0, as the base model. Qwen3.5 is by the Qwen team, Alibaba Cloud.