OSS-forge/DeepSeek-Coder-1.3B-cleaned

DeepSeek-Coder-1.3B – Clean DSC Model (DSCc)

This repository hosts DSCc, a fine-tuned version of DeepSeek-Coder-1.3B trained for Python function generation from docstrings and function signatures, using a cleaned subset of The Stack.

The model is part of the study:

Quality In, Quality Out: Investigating Training Data’s Role in AI Code Generation
33rd IEEE/ACM International Conference on Program Comprehension (ICPC 2025)

DSCc is specifically trained on a Semgrep-filtered dataset that removes many low-quality and syntactically incorrect functions, allowing us to study how training data quality impacts code generation performance.

---

Model description

• Base model: DeepSeek-Coder-1.3B (Python-focused code LLM)
• Task: Python code generation
• Input: Python function docstring + signature
• Output: The corresponding function body in Python

In our experiments, the model is conditioned on a prompt consisting of:

• A natural-language docstring describing the function behavior
• The Python function signature

and is then asked to generate the rest of the function body.

---

What does the model do?

The model generates Python functions that implement the behavior described in the docstring and implied by the signature. Typical use cases:

• Synthesizing a function implementation from a high-level description
• Suggesting implementations for partially specified functions
• Exploring how training data quality affects generated code (correctness, style, quality issues)

“Clean” training set (for DSCc)

The initial training set contains ~4.4M pairs. To construct the clean dataset:

• We run Semgrep (static analysis) on all training functions.
• Semgrep detects:
  • Low-quality patterns
  • Potentially problematic constructs
  • Syntactically incorrect functions
• All flagged low-quality / invalid functions are removed.

This yields:

• clean_training_set.json — ~4.2M pairs
  • Derived from The Stack
  • But with many quality issues and syntax errors filtered out.