| --- |
| license: apache-2.0 |
| language: |
| - en |
| - pt |
| - zh |
| - es |
| - de |
| - it |
| - fr |
| metrics: |
| - type: |
| HumanEval zero-shot pass@1: null |
| value: |
| '88.41': null |
| 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 <u>partially uncensored</u> model focused in <b> coding</b>. 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) |
| <br> |
| <p> |
|
|
| #### Nerdsking Python Coder Family |
|
|
| 🧠 <a href="https://huggingface.co/Nerdsking/nerdsking-python-coder-3B-i"> Nerdsking Python Coder 3B-i </a><br> |
| 🧠 <a href="https://huggingface.co/Nerdsking/Nerdsking-python-coder-7B-i"> Nerdsking Python Coder 7B-i </a> |
| <br> |
| <p> |
|
|
|
|
| #### 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 completely 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> |
| |
| <hr> |
| <br><b>IMPORTANT:</b> 5 "errors" from the model during the benchmark were mere "import errors" (missing imports: reduce, Optional, List, etc ), but the logic was perfect, therefore: |
| - the model reasoning was right |
| - the failure is syntactic / boilerplate, not conceptual<br> |
| |
| We did not considered it for our score, but "if" considered those extra 5 questions as correct, our benchmark would be <b>much higher</b>. |
| <hr> |
| |
| |
| #### Comparison Table with models of the same size |
| |
| <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>Gemma 3 27B</td> |
| <td>76</td> |
| <td>Our model surpasses a model 9x bigger by 12%</td> |
| </tr> |
| <tr> |
| <td>Qwen/Qwen3-4B-GGUF (<u>bigger</u> size model)</td> |
| <td>~73%</td> |
| <td>Indicative proxy from published code-task performance breakdowns (not a strict HumanEval pass@1)</td> |
| </tr> |
| <tr> |
| <td>Llama 3.3 70B</td> |
| <td>~83% </td> |
| <td>Our model surpasses a model 23x bigger</td> |
| </tr> |
| <tr> |
| <td>CodeLlama 7B‑Python (<u>much bigger</u> size model)</td> |
| <td>~74 % </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> |
| |
| |
| <hr> |
| |
| ### TIME Comparison Table with models of Bigger size WITH "thinking mode" |
| |
| |
| ### <font color=red>Nanbeige/Nanbeige4.1-3B </font> (4 billion model, "marketed" as 3B billions), with 92% HumanEval accuracy in Python (FP16) |
| - Time to finish 160 questions from HumanEval: 22263.80s --- <b><u>MORE THAN SIX HOURS</b></u> !!! (in a RTX 5060 ti 16gb) |
| |
| ### <font color=red>Nerdsking/nerdsking-python-coder-3B-i</font> (real 3 billion model) with 88,41% HumanEval accuracy in Python (FP16) |
| - Time to finish 160 questions from HumanEval: 210.51s --- <b><u>THREE MINUTES AND HALF</b></u> !!! (in a RTX 5060 ti 16gb) |
| |
| |
| <hr> |
| |
| |
| |
| #### Benchmark tool used |
| |
| https://github.com/nerdskingcom/gguf-humaneval-benchmark |
| |
| Install it using: |
| |
| <code> |
| pip install gguf-humaneval-benchmark |
| </code> |
| |
| Instructions after install: |
| |
| <code> |
| gguf-humaneval-benchmark --help |
| </code> |
| |
| <hr> |
| |
| |
| |
| #### S.o.n.n. |
| <p class="justified-text"> |
| The model was treated under <b>"s.o.n.n."</b> (<i>singular 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/or 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, |
| Iran Necho (IPMN) / Nerdsking.com |
