Nerdsking/Nerdsking-python-coder-7B-i

1---2license: apache-2.03language:4- en5- pt6- zh7- es8- de9- it10- fr11metrics:12- type:13    HumanEval zero-shot pass@1: null14  value:15    '86.99': null16pipeline_tag: text-generation17tags:18- code19---20 21 22<!-- Provide a quick summary of what the model is/does. -->23 24#### Model Details25<p class="justified-text">26<b>Nerdsking-python-coder-7B-i</b> is a 7B parameter partially uncensored 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.27</p>28<i>Key Characteristics:</i>29 30- Parameter count: 7B31- Primary domain: Python programming32- Secondary capabilities: General coding, technical English33- Training focus: Python logic, standard library usage, algorithmic reasoning34- Alignment: Partially uncensored (developer-oriented)35<br>36<p>37 38#### Nerdsking Python Coder Family39 40🧠 <a href="https://huggingface.co/Nerdsking/nerdsking-python-coder-3B-i"> Nerdsking Python Coder 3B-i </a><br>41🧠 <a href="https://huggingface.co/Nerdsking/Nerdsking-python-coder-7B-i"> Nerdsking Python Coder 7B-i </a>42<br>43<p>44  45#### Benchmark46<p class="justified-text">47After intense refining, <b>Nerdsking-python-coder-7B-i</b> has achieved <b>86.99 in HumanEval (bf16)</b>, ranking it amongst the highest-performing Python-focused 7B models ever reported on HumanEval. Surpassing even much bigger models in that area. 48</p>49<i>Benchmark details (164 tasks):</i>50 51- official HumanEval execution protocol - test suites executed via `exec()`52- zero-shot pass@153- dtype == "bfloat16"54- temperature = 0.155- do_sample = False56- evaluated on fully merged weights57- Prompting: Chat-formatted with a fixed system prompt (“You are an expert Python coding assistant.”)58- Quantization: None (unquantized weights - bf16)59<p class="justified-text">60<i>The configuration above is fully disclosed to support reproducibility and fair comparison.</i>61</p>62<p class="justified-text">63<i> Note: Quantized variants (INT4/INT6) may exhibit lower HumanEval scores due to reduced numerical precision.</i>64</p>65 66 67#### Comparison Table68 69<table>70  <thead>71    <tr>72      <th>Model name</th>73      <th>Approx. HumanEval Pass@1 (%)</th>74      <th>Notes / Source</th>75    </tr>76  </thead>77  <tbody>78    <tr>79      <td><strong>Nerdsking-python-coder-7B-i</strong></td>80      <td><strong>86.99</strong></td>81      <td>Evaluated score (zero-shot, strict HumanEval pass@1, using unquantized weigths bf16)</td>82    </tr>83    <tr>84      <td>Qwen2.5-Coder-7B</td>85      <td>~74–76</td>86      <td>Community evaluation (OpenCompass run); figures vary by harness/settings</td>87    </tr>88    <tr>89      <td>DeepSeek-Coder-6.7B</td>90      <td>~72–73</td>91      <td>Official DeepSeek report and independent replications; close to strict HumanEval protocol</td>92    </tr>93    <tr>94      <td>CodeLlama-7B</td>95      <td>~33–35</td>96      <td>Meta technical report</td>97    </tr>98    <tr>99      <td>Wizard Coder 7B*</td>100      <td>~57–59</td>101      <td>Community benchmarks; strong instruction-following but less consistent zero-shot behavior</td>102    </tr>103  </tbody>104</table>105<p class="justified-text">106  107</p>108 109<hr>110 111 112#### Benchmark tool used113 114https://github.com/nerdskingcom/gguf-humaneval-benchmark115 116Install it using:117 118<code>119 pip install gguf-humaneval-benchmark120</code>121 122Instructions after install:123 124<code>125 gguf-humaneval-benchmark --help126</code>127 128<hr>129 130 131#### S.o.n.n.132<p class="justified-text">133The 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.134</p>135<i>When applied to pre-existing models, allows:</i>136 137- parameter-preserving refinement methodology138- focused global behavioral shaping, instead of task-local adapters139- avoidance of fragmentation, very common in multi-adapter or task-siloed approaches140 141 142 143#### Quick Start (Inference)144 145<code>146from transformers import AutoModelForCausalLM, AutoTokenizer147 148model_id = "Nerdsking/Nerdsking-python-coder-7B-i"149 150tokenizer = AutoTokenizer.from_pretrained(model_id)151model = AutoModelForCausalLM.from_pretrained(152    model_id,153    torch_dtype="bfloat16",154    device_map="auto"155)156 157prompt = "Write a Python function that checks if a number is prime."158inputs = tokenizer(prompt, return_tensors="pt").to(model.device)159outputs = model.generate(**inputs, max_new_tokens=200)160 161print(tokenizer.decode(outputs[0], skip_special_tokens=True))162</code>163 164#### Ethical & Safety Notes165<p class="justified-text">166This model is intended for technical and research use.167Due to relaxed alignment constraints, outputs should be reviewed before deployment in production or public-facing systems.168</p>169 170#### Citation171 172If you use this model in research and/or benchmarking, please cite:173 174Nerdsking-python-coder-7B-i,175Iran Necho (IPMN) / Nerdsking.com