README.md

---
license: fair-noncommercial-research-license
language:
- en
- pt
metrics:
      - type: { HumanEval zero-shot pass@1}         # Required. Example: wer. Use metric id from https://hf.co/metrics
        value: {88.41}       # Required. Example: 20.90
base_model:
- Qwen/Qwen2.5-Coder-3B-Instruct
pipeline_tag: text-generation
tags:
- code
---


<!-- Provide a quick summary of what the model is/does. -->

#### Model Details
<p class="justified-text">
<b>Nerdsking-python-coder-3B-i</b> is a 3B parameter partially uncensored model focused in <b> Python</b>, with <b>English</b> as main language. It was massively trained in python, therefore despite the fact it can code in other languages as well, the performance will be not in the same level as the one achieved while using python.
</p>
<i>Key Characteristics:</i>

- Parameter count: 3B
- Primary domain: Python programming
- Secondary capabilities: General coding, technical English
- Training focus: Python logic, standard library usage, algorithmic reasoning
- Alignment: Partially uncensored (developer-oriented)


#### Benchmark
<p class="justified-text">
After intense refining, <b>Nerdsking-python-coder-3B-i</b> has achieved <b>88.41 in HumanEval (bf16)</b>, ranking it amongst the highest-performing Python-focused 3B models ever reported on HumanEval. Surpassing even much bigger models in that area. 
</p>
<i>Benchmark details (164 tasks):</i>

- official HumanEval execution protocol - test suites executed via `exec()`
- zero-shot pass@1
- dtype == "bfloat16"
- temperature = 0.1
- do_sample = False
- evaluated on fully merged weights
- Prompting: Chat-formatted with a fixed system prompt (“You are an expert Python coding assistant.”)
- Quantization: None (unquantized weights - bf16)
<p class="justified-text">
<i>The configuration above is fully disclosed to support reproducibility and fair comparison.</i>
</p>
<p class="justified-text">
<i> Note: Quantized variants (INT4/INT6) may exhibit lower HumanEval scores due to reduced numerical precision.</i>
</p>


#### Comparison Table

<table>
  <thead>
    <tr>
      <th>Model name</th>
      <th>Approx. HumanEval Pass@1 (%)</th>
      <th>Notes / Source</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td><strong>Nerdsking-python-coder-3B-i</strong></td>
      <td><strong>88.41</strong></td>
      <td>Evaluated score (zero-shot, strict HumanEval pass@1, using unquantized weigths bf16)</td>
    </tr>
    <tr>
      <td>Qwen2.5-3B-Instruct</td>
      <td>~45 (community)</td>
      <td>Community evaluation (OpenCompass run); figures vary by harness/settings</td>
    </tr>
    <tr>
      <td>StarCoder2-3B</td>
      <td>~33.6</td>
      <td>Reported in third-party performance overview; may differ by protocol</td>
    </tr>
    <tr>
      <td>Stable Code 3B*</td>
      <td>~32–33 (estimate)</td>
      <td>Indicative proxy from published code-task performance breakdowns (not a strict HumanEval pass@1)</td>
    </tr>
    <tr>
      <td>Wizard Coder 3B*</td>
      <td>~31.6 (estimate)</td>
      <td>Indicative proxy from published code-task performance breakdowns (not a strict HumanEval pass@1)</td>
    </tr>
    <tr>
      <td>StarCoder 3B*</td>
      <td>~21.6 (estimate)</td>
      <td>Indicative proxy from published code-task performance breakdowns (not a strict HumanEval pass@1)</td>
    </tr>
  </tbody>
</table>
<p class="justified-text">
  <em>*Estimated/proxy values where standardized HumanEval pass@1 was not published in those 3 models. Scores can vary with prompt format, decoding params, and harness.</em>
</p>




#### S.o.n.n.
<p class="justified-text">
The model was treated under <b>"s.o.n.n."</b> (<i>single omni neural network</i>), a concept created by IPMN at Nerdsking.com that is both a precise way of fine tunning/altering existing models, as well a foundational concept for a broader AI architecture standard currently under active research and development.
</p>
<i>When applied to pre-existing models, allows:</i>

- parameter-preserving refinement methodology
- focused global behavioral shaping, instead of task-local adapters
- avoidance of fragmentation, common in multi-adapter or task-siloed approaches



#### Quick Start (Inference)

<code>
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Nerdsking/Nerdsking-python-coder-3B-i"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="bfloat16",
    device_map="auto"
)

prompt = "Write a Python function that checks if a number is prime."
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=200)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))
</code>

#### Ethical & Safety Notes
<p class="justified-text">
This model is intended for technical and research use.
Due to relaxed alignment constraints, outputs should be reviewed before deployment in production or public-facing systems.
</p>

#### Citation

If you use this model in research or benchmarking, please cite:

Nerdsking-python-coder-3B-i,
IPMN / Nerdsking.com