Base Model
Base model: Qwen/Qwen2.5-Coder-3B-Instruct.
This is a 3B base model. The point of this release is to start building small, task-specialized coding ELMs that can become smarter per parameter than broad general-purpose LLMs for focused engineering workflows.
Training
Training used a manually curated coding mixture built from public coding resources and private coding/AI-engineering work sessions that were cleaned before training. The raw private dataset is not published. The release includes proof artifacts rather than raw source data.
This adapter is the selected DPO-stage adapter from the proof-gated run.
Proof
Executable benchmark: 100 HumanEval tasks on Kaggle GPU.
Table with columns: Evaluation, Pass count| Evaluation | Pass count |
|---|
| Base model | 65 / 100 |
| After SFT | 68 / 100 |
| After DPO | 68 / 100 |
| Final selected model | 68 / 100 |
Absolute lift over base: +3.0 percentage points on the 100-task HumanEval proof run.
Proof files are included in this repo: eval_before_after_full_code.csv, release_summary.json, trainer_log_history.json, and related partial eval CSVs.
Use
Load this adapter on top of Qwen/Qwen2.5-Coder-3B-Instruct with PEFT. For normal use without adapter loading, use the merged model repo.
Evidence files
Run evidence for this release is stored in the repository under evidence/:
These files are compact local/Kaggle run artifacts used to document training, evaluation, merge, or quantization evidence for this model family.