OpenSonnet-Lite-MAX/README.md

---
base_model:
  - Qwen/Qwen3-4B-Thinking-2507

datasets:
  - Roman1111111/claude-sonnet-4.6-120000x
  - Roman1111111/claude-sonnet-4.6-100000X-filtered
  - TeichAI/lordx64-claude-opus-4.7-max-cleaned
  - Crownelius/Opus-4.6-Reasoning-3300x
  - TeichAI/claude-4.5-opus-high-reasoning-250x
  - TeichAI/claude-haiku-4.5-high-reasoning-1700x
  - TeichAI/claude-sonnet-4.5-high-reasoning-250x
  - TeichAI/deepseek-v3.2-speciale-openr1-math-3k
  - TeichAI/deepseek-v3.2-speciale-1000x
  - Roman1111111/gemini-3-pro-10000x-hard-high-reasoning
  - Roman1111111/gemini-3.1-pro-hard-high-reasoning
  - Jackrong/DeepSeek-V4-Distill-8000x

tags:
  - opensonnet
  - claude-sonnet
  - sonnet

pipeline_tag: text-generation
library_name: transformers
license: apache-2.0
license_link: https://huggingface.co/hadadxyz/OpenSonnet-Lite-MAX/blob/main/LICENSE
---

# Comparison

| Model                                                                        | Training Approach        | Developer Role         | Context Length | Training Epochs  | Transformers Version   | Notes                                                                               |
|------------------------------------------------------------------------------|--------------------------|------------------------|----------------|------------------|------------------------|-------------------------------------------------------------------------------------|
| [OpenSonnet-Lite-MAX](https://huggingface.co/hadadxyz/OpenSonnet-Lite-MAX)   | Multi-Stage Fine-Tuning  | Supported              | 262,144        | 2                | `transformers>=5.0.0`  | Latest version with improved training efficiency and enhanced instruction alignment |
| [OpenSonnet-Lite](https://huggingface.co/hadadxyz/OpenSonnet-Lite)           | Single-Stage Fine-Tuning | Not supported          | 262,144        | 3                | `transformers>=4.51.0` | Previous version with simpler training pipeline                                     |
| [Qwen3-4B-Thinking-2507](https://huggingface.co/Qwen/Qwen3-4B-Thinking-2507) | N/A                      | Not supported          | 262,144        | N/A              | `transformers>=4.51.0` | Base model                                                                          |

> [OpenSonnet-Lite-MAX quick demo](https://www.kaggle.com/code/hadadrjt/opensonnet-lite-max) with tool calling.

### Benchmark Evaluation

| Dataset                                               | Score  | Source                                                                                            | Framework                                                                    |
|-------------------------------------------------------|--------|---------------------------------------------------------------------------------------------------|------------------------------------------------------------------------------|
| [GSM8K](https://huggingface.co/datasets/openai/gsm8k) | 85.22  | [Evaluation Results](https://huggingface.co/hadadxyz/OpenSonnet-Lite-MAX/tree/main/.eval_results) | [lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness) |
| MMLU-Pro                                              | -      | -                                                                                                 | -                                                                            |
| GPQA (Diamond)                                        | -      | -                                                                                                 | -                                                                            |


# Inference Parameters

For best results, the following sampling configuration is recommended:

| Parameter           | Recommended Value   | Description                              |
|---------------------|---------------------|------------------------------------------|
| temperature         | 0.6 (default) - 1.0 | Controls randomness in generation        |
| top_p               | 0.95 (default)      | Nucleus sampling threshold               |
| top_k               | 20 (default) - 40   | Top-k sampling parameter                 |
| min_p               | 0.0 (default)       | Minimum probability threshold            |
| repetition_penalty  | 1.0 (default) - 1.2 | Penalizes repeated tokens                |
| presence_penalty    | 1.0 - 1.5           | Encourages introducing new topics        |


# Max Tokens

| Small Tasks | Medium Tasks | Large Tasks | Complex Tasks |
|-------------|--------------|-------------|---------------|
| 4096/8192   | 16384        | 32768/81920 | 131072        |


# Instruction

```md
You are OpenSonnet, a large language model trained by the Open Source community. You are based on the Qwen3 architecture.

You are an AI assistant designed to provide accurate, helpful, and context-aware responses. Your reasoning style must dynamically adapt based on the complexity of the user’s request.

---

# Adaptive Thinking Mode

* Automatically assess the complexity of each user request before responding.

* If the task is complex, multi-step, analytical, or requires planning, reasoning, or explanation:
  - Use structured, step-by-step reasoning internally before responding.
  - Provide a clear, well-organized, and thorough answer.

* If the task is simple, factual, or straightforward:
  - Use fast, minimal reasoning.
  - Respond concisely without unnecessary elaboration.

---

# Complexity Detection Guidelines

* Treat a request as COMPLEX if it involves:
  - Multi-step problem solving
  - Logic, mathematics, coding, or debugging
  - Planning, strategy, or decision making
  - Deep explanation or comparison
  - Ambiguous or multi-part instructions

* Treat a request as SIMPLE if it involves:
  - Direct factual questions
  - Basic definitions
  - Short instructions
  - Common knowledge retrieval
  - Single-step tasks

---

# Response Style Rules

* Always prioritize correctness and clarity.

* For complex tasks: structure answers clearly using sections or bullet points when helpful.

* For simple tasks: keep responses short and direct.

* Avoid unnecessary verbosity in all cases.

---

# Quality Principles

* Be accurate, logical, and consistent.

* Do not hallucinate information.

* If uncertain, clearly state limitations.

* Optimize responses for usefulness and readability.

---

# User Intent Focus

* Always prioritize the user’s intent over literal interpretation.

* If the request is ambiguous, make reasonable assumptions or ask a clarifying question when necessary.
```


# Citation

If you use this model in your research or applications, please cite both this model and the base model:

```bibtex
@misc{opensonnet-lite-max,
  author = {hadadxyz},
  title  = {OpenSonnet-Lite-MAX},
  year   = {2026},
  url    = {https://huggingface.co/hadadxyz/OpenSonnet-Lite-MAX}
}
```