maxx-1-1.5B/README.md

---
language:
- en
license: apache-2.0
base_model: Qwen/Qwen2.5-1.5B-Instruct
tags:
- qwen2
- unsloth
- trl
- lora
- on-device
- agentic
- offline
- fine-tuned
model_type: qwen2
pipeline_tag: text-generation
---

# maxx — On-Device Agentic LLM (1.5B)

> A fine-tuned Qwen2.5-1.5B-Instruct model optimized for agentic tasks,
> instruction following, and real-world offline use on phones and laptops.
> First checkpoint in an ongoing research project targeting the best
> open-source agentic model under 3B parameters.

---

## Model Details

| Field | Details |
|---|---|
| **Base model** | Qwen/Qwen2.5-1.5B-Instruct |
| **Parameters** | 1.5B |
| **Fine-tune method** | QLoRA (4-bit, rank 16) |
| **Framework** | Unsloth + TRL |
| **Context window** | 2048 tokens |
| **License** | Apache 2.0 |
| **Developer** | bolajiev (Independent Researcher) |
| **Status** | EXP-001 — active research |

---

## Benchmark Results (EXP-001)

Evaluated using [lm-evaluation-harness](https://github.com/EleutherAI/lm-harness) with 5-shot prompting.

| Benchmark | maxx (1.5B) | Qwen2.5-1.5B-Instruct | SmolLM2-1.7B-Instruct |
|---|---|---|---|
| ARC-Challenge ↑ | 52.47% | **53.92%** | 51.88% |
| HellaSwag ↑ | 67.02% | 67.71% | **72.20%** |
| WinoGrande ↑ | 65.51% | 64.64% | **68.98%** |
| TruthfulQA ↑ | 45.99% | 46.61% | 39.96% |
| **MMLU ↑** | **59.87%** | — | — |
| **Average** | **57.75%** | 58.22% | 58.26% |

**Key findings:**
- Within **0.5%** of both larger/better-resourced competitors on first training run
- Beats SmolLM2-1.7B on TruthfulQA by **+6 points** — a bigger model
- MMLU of **59.87%** outperforms published reference scores for both competitors
- Strong commonsense and knowledge base retained from Qwen2.5 foundation

---

## Intended Use

### Primary use cases
- On-device AI assistant for phones and laptops (no internet required)
- Instruction following and task completion offline
- Summarization, email writing, scheduling, planning
- Agentic multi-step reasoning for everyday tasks
- Privacy-first AI — all compute runs locally

### Out of scope
- High-stakes medical, legal, or financial decisions
- Tasks requiring real-time internet access
- Complex multi-modal tasks

---

## Training Details

### Data
- OpenHermes-2.5 — instruction following
- UltraChat-200k — conversational quality
- Glaive Function Calling v2 — tool use and agentic tasks
- Alpaca Cleaned — general instructions
- Synthetic data generated via open-source teacher model (Qwen2.5-7B)

**Total:** ~35,000 curated examples (EXP-001 small run)

### Hyperparameters

| Parameter | Value |
|---|---|
| Learning rate | 2e-4 |
| Batch size | 4 |
| Gradient accumulation | 4 (effective 16) |
| LoRA rank | 16 |
| LoRA alpha | 32 |
| Max steps | 200 |
| Optimizer | AdamW 8-bit |
| Scheduler | Cosine |
| Warmup steps | 20 |

### Hardware
- GPU: Kaggle T4 (16GB VRAM)
- Training time: ~1.5 hours
- Compute: ~3 GPU hours

---

## How to Use

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "bolajiev/maxx-1-1.5B"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
    device_map="auto",
)

messages = [{"role": "user", "content": "Write a short email to my boss saying I will be 10 minutes late."}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

with torch.no_grad():
    output = model.generate(**inputs, max_new_tokens=300, temperature=0.7, do_sample=True)

reply = tokenizer.decode(output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
print(reply)
```

### On-Device with Ollama (GGUF)
```bash
# Use the quantized GGUF version for on-device inference
ollama run bolajiev/maxx-merged-gguf
```

---

## Limitations

- EXP-001 is a small training run (200 steps, ~35k examples) — not a final model
- Safety alignment is limited — some harmful requests may not be refused correctly
- Context window limited to 2048 tokens in this checkpoint
- Not evaluated on coding tasks yet
- HellaSwag gap vs SmolLM2 indicates commonsense reasoning can improve

---

---


---

*Built with [Unsloth](https://github.com/unslothai/unsloth) 🦥 | Trained on Kaggle T4*