Executespec/ganesh-python-v0.1.0

1---2language:3  - en4library_name: transformers5pipeline_tag: text-generation6tags:7  - executespec-rd-lab8  - code9  - coding10  - python11  - safetensors12  - gguf13  - gemma14base_model: google/gemma-4-12B-it15license: gemma16---17 18# Ganesh Python v0.1.019 20Ganesh Python v0.1.0 is a 12B coding model focused on Python code generation,21code repair and developer-assistant workflows. It is distributed as a22standalone merged checkpoint based on Gemma 4 12B IT. It is developed and23released by the ExecuteSpec R&D Lab.24 25**Project status (2026-08-27):** supported, frozen research release. A later26unified v0.2 experiment failed its executable specialist-preservation gate and27was not released. Model development is closed; this v0.1 specialist remains28the final published Ganesh Python artifact.29 30## Intended use31 32- Python code generation and completion33- Code repair and debugging assistance34- Refactoring and implementation exploration35- Coding-agent and developer-tool evaluation36- Research and experimentation with local or hosted inference37 38## Model format39 40- Architecture family: Gemma 4 12B IT41- Weight formats: merged BF16 SafeTensors, GGUF Q8_0 and GGUF Q4_K_M42- Interface: text generation43- Version: v0.1.044 45This repository contains a complete standalone model; no additional weight46package is required for inference.47 48## Usage49 50Use the tokenizer and processor files included in this repository. Runtime51examples will be added after clean-environment compatibility validation for52Transformers, vLLM and Unsloth.53 54For local llama.cpp inference, select a file from `gguf/`. Q8_0 prioritizes55fidelity; Q4_K_M reduces memory and storage requirements.56 57## Evaluation snapshot58 59On the retained Python executable evaluation, the base scored 16/100 and this60specialist scored 18/100 on development. On the sealed 200-task decision set,61the base scored 14/200 and this specialist scored 23/200. On EvalPlus, the62reported base-to-specialist results were HumanEval 95.1→96.3, HumanEval+6392.1→93.9, MBPP 87.8→88.4, and MBPP+ 73.3→73.8.64 65These are narrow, versioned evaluation surfaces, not a claim of general coding66superiority. HumanEval-family saturation and historical data availability, and67MBPP training adjacency, materially limit interpretation.68 69## Limitations70 71This is an early coding-model release intended for evaluation. It may produce72incorrect, incomplete, insecure or inefficient code; invent APIs; mishandle73edge cases; or fail to follow repository-specific conventions. Generated code74must be reviewed and tested before use. Do not rely on the model for75security-critical, safety-critical or compliance-sensitive decisions without76independent verification.77 78Performance may vary with prompt format, runtime, precision, sampling settings,79context length and task distribution. Results from one runtime should not be80assumed to transfer unchanged to another.81 82## Versioning83 84Ganesh v0.1.0 publishes separate language-focused model identities. No unified85Ganesh successor was released. This repository is frozen except for factual,86safety, licensing, and documentation corrections.87 88## Attribution89 90Ganesh Python v0.1.0 is developed and released by the ExecuteSpec R&D Lab and91is based on Gemma 4 12B IT. Use of this model remains subject to the applicable92Gemma Terms of Use and Gemma Prohibited Use Policy.93