README.md

# Inelly 4.5 Blaze

## Model Description

**Inelly 4.5 Blaze** is a fine-tuned version of [Qwen2.5-1.5B-Instruct](https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct), trained on a focused mixture of chain-of-thought reasoning, math, coding, and general knowledge data. It is the compact, fast variant of the Inelly 4.5 family -- optimized for quick inference while retaining strong reasoning capabilities.

- **Developed by:** bry 
- **Base model:** Qwen2.5-1.5B-Instruct
- **Fine-tuning method:** QLoRA (4-bit NF4, rank 16)
- **Parameters:** 1.54B (base) + ~3.1M trainable (LoRA adapters)
- **License:** Apache 2.0 (inherited from Qwen2.5)

---

## Intended Use

Inelly 4.5 Blaze is intended for:

- **Chain-of-Thought reasoning** – Step-by-step problem solving
- **Math** – Algebra, arithmetic, word problems
- **Code generation** – Python functions with clear logic
- **Logical deduction** – Syllogisms, puzzles, multi-step reasoning
- **General knowledge Q&A** – Science, everyday facts
- **Quick prototyping** – Fast inference on consumer hardware

### Out of Scope

- Not intended for production deployment without further safety evaluation
- Less conversational polish than the 3B variant (Inelly 4.5)
- May struggle with very long or complex multi-step tasks

---

## Training Data

Inelly 4.5 Blaze was fine-tuned for 1 epoch on ~5,225 samples drawn from:

| Dataset | Samples | Purpose |
|---|---|---|
| [Bespoke-Stratos-35k](https://huggingface.co/datasets/bespokelabs/Bespoke-Stratos-35k) | 3,000 | Chain-of-thought math & reasoning |
| [OpenThoughts-114k](https://huggingface.co/datasets/open-thoughts/OpenThoughts-114k) | 2,500 | Code generation with reasoning |
| [dolphin-r1](https://huggingface.co/datasets/cognitivecomputations/dolphin-r1) | 2,000 | General reasoning (DeepSeek-R1 distill) |

All samples were deduplicated and reasoning-weighted (2x oversample for CoT examples). Maximum sequence length: 512 tokens.

---

## Training Hyperparameters

| Parameter | Value |
|---|---|
| Base model | Qwen2.5-1.5B-Instruct |
| Quantization | 4-bit NF4 (bitsandbytes) |
| LoRA rank | 16 |
| LoRA alpha | 32 |
| LoRA dropout | 0.05 |
| Learning rate | 2e-4 |
| Batch size | 8 (gradient accumulation) |
| Epochs | 1 |
| Max seq length | 512 |
| Optimizer | AdamW 8-bit |
| LR scheduler | cosine |
| Warmup ratio | 0.05 |
| Training time | ~35 min |
| Hardware | RTX 2080 Ti (11GB VRAM) |

---

## Model Architecture

| Property | Value |
|---|---|
| Model type | Qwen2ForCausalLM |
| Hidden size | 1,536 |
| Layers | 28 |
| Attention heads | 12 |
| Head dim | 128 |
| Intermediate size | 8,960 |
| Vocab size | 151,936 |
| Context length | 32,768 |
| Total parameters | ~1.54B |
| Trainable parameters | ~3.1M (LoRA) |

---

## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("path/to/inelly-4.5-blaze", torch_dtype=torch.float16, device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("path/to/inelly-4.5-blaze")

messages = [{"role": "user", "content": "Solve for x: 3x + 7 = 22. Show all steps."}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

output = model.generate(**inputs, max_new_tokens=256, temperature=0.7, top_p=0.9)
response = tokenizer.decode(output[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
print(response)
```

---

## Performance

Informal GPU testing across 8 categories:

| Category | Result |
|---|---|
| Chain-of-Thought reasoning | ✅ Correct step-by-step logic |
| Math | ✅ Accurate algebraic solutions |
| Code generation | ✅ Clean Python with comments |
| Logic puzzles | ✅ Sound deductive reasoning |
| General knowledge | ✅ Accurate, clear explanations |
| Speed | ✅ ~1-2s per response (faster than 3B/7B) |

---

## Inelly 4.5 Family Comparison

| Model | Size | Focus | Training Data |
|---|---|---|---|
| **Inelly 4.5** | 3B | Conversation + CoT | 5,700 samples (incl. politeness, conv) |
| **Inelly 4.5 Blaze** (this) | 1.5B | Fast reasoning + CoT | 5,225 samples (reasoning-focused) |
| Matrix 2| 7B | Deep reasoning | 5,225 samples (reasoning-focused) |

**When to use Blaze vs standard 4.5:**
- **Blaze** – When you need fast reasoning, math, or coding help and don't need conversational polish
- **4.5 (3B)** – When you want a friendly, polite conversationalist that can also reason

---

## Limitations

- **Conversational ability:** Less polished in casual chat compared to the 3B variant (no conversational fine-tuning data)
- **Safety:** Inherited from Qwen2.5 base; not specifically safety-tuned
- **Context length:** Fine-tuned on 512-token sequences
- **Factual accuracy:** May hallucinate in specialized domains

---

## Acknowledgments

- [Qwen2.5](https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct) by Alibaba Cloud (base model)
- [Bespoke Labs](https://huggingface.co/bespokelabs) for Stratos dataset
- [OpenThoughts](https://huggingface.co/datasets/open-thoughts/OpenThoughts-114k) team
- [Cognitive Computations](https://huggingface.co/cognitivecomputations) for dolphin-r1

---

## Citation

```
@misc{inelly45blaze,
  title = {Inelly 4.5 Blaze: A Compact Chain-of-Thought Reasoning Model},
  author = {Genue},
  year = {2026},
  note = {Fine-tuned from Qwen2.5-1.5B-Instruct using QLoRA},
}
```