---
license: apache-2.0
datasets:
- OpenAssistant/oasst1
---

# 🍌 BananaMind V1

> ⚠️ **Honest disclaimer:** This model is not very good. It's a small experimental model trained as a personal project. Manage your expectations accordingly.

BananaMind V1 is a 125M parameter causal language model fine-tuned on conversational data. It is uploaded in **GGUF format** for use with llama.cpp and compatible runtimes.

---

## Model Details

| Property | Value |
|---|---|
| **Model Name** | BananaMind V1 |
| **Parameters** | 125M |
| **Context Length** | 512 tokens |
| **Format** | GGUF |
| **License** | Apache 2.0 |
| **Language** | English |

---

## Training Data

| Dataset | License |
|---|---|
| [OpenAssistant/oasst1](https://huggingface.co/datasets/OpenAssistant/oasst1) | Apache 2.0 |

---

## Intended Use

- Experimenting with small language models
- Local inference on CPU via llama.cpp
- Educational / research purposes
- Tinkering and testing

**Not intended for:** Production use, safety-critical applications, or anything requiring reliable or accurate outputs.

---

## How to Use

### With llama.cpp

```bash
./llama-cli -m BananaMind-V1.gguf \
  -p "User: Hello! How are you?\nAssistant:" \
  -n 128 \
  --ctx-size 512
```

### With Python (via llama-cpp-python)

```python
from llama_cpp import Llama

llm = Llama(
    model_path="BananaMind-V1.gguf",
    n_ctx=512,
)

output = llm(
    "User: What is the capital of France?\nAssistant:",
    max_tokens=128,
    stop=["User:"],
)

print(output["choices"][0]["text"])
```

### With Ollama
```bash
ollama run banaxitech/bananamind-v1
``


---

## Limitations & Known Issues

- **Small capacity:** 125M parameters is quite limited — don't expect GPT-4 quality.
- **Short context:** 512 token context window means it forgets things quickly.
- **Hallucinations:** Will confidently make things up.
- **Inconsistency:** Outputs can vary wildly between runs.
- **Generally not very good** — as stated upfront. Consider this a starting point, not a destination.

---

## Evaluation

No formal benchmarks were run. Informal testing suggests the model can hold basic conversations but struggles with anything requiring reasoning, factual recall, or multi-turn coherence.

---

## License

This model is released under the **Apache 2.0** license, consistent with its training data. See [LICENSE](https://www.apache.org/licenses/LICENSE-2.0) for details.

---

## Citation

If you somehow use this model in research (why?), you can cite it informally as:

```
BananaMind V1 (2024). A small 125M parameter GGUF language model. Personal project.
```

---

*Made with 🍌 and questionable judgment.*