GX-Coder-7B

v0.1 — QLoRA fine-tune of unsloth/Qwen2.5-Coder-7B-Instruct, built to power a multi-mode coding agent. This first release sits at parity with the base on HumanEval+ (see table); it's an end-to-end QLoRA→eval→ship baseline. Later versions target agent-specific tasks (tool-calling, web/UI codegen) where the base isn't already saturated.

Summary

• Base: unsloth/Qwen2.5-Coder-7B-Instruct
• Method: 4-bit QLoRA (LoRA r=16) via Unsloth
• Hardware: single free-Colab T4 (16GB)
• Data: open instruction-coding datasets, ChatML-formatted (see training repo)

Benchmarks (HumanEval+ pass@1, greedy)

Model	HumanEval+ pass@1
unsloth/Qwen2.5-Coder-7B-Instruct (base)	80.5
GX-Coder-7B (this)	78.0

Scored with EvalPlus (164 problems, greedy). The 2pt gap is within noise (3-4 problems) — treat as parity. A generic fine-tune can't easily beat a near-ceiling base here; gains come from task-specific data in later versions.

Intended use

Code generation, completion, review, and tool-using agent workflows. Powers the companion multi-mode coding agent (Claude-Code-like coder + Figma-like web/UI design mode).

Limitations

Small model fine-tune — not frontier-level. May hallucinate APIs, miss edge cases, and produce insecure code. Always review generated code before running. Not trained on benchmark test sets (contamination guard in data_prep.py).

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("Garv98/GX-Coder-7B")
model = AutoModelForCausalLM.from_pretrained("Garv98/GX-Coder-7B", device_map="auto")
msgs = [{"role": "user", "content": "Write a Python function to check if a string is a palindrome."}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(model.device)
print(tok.decode(model.generate(ids, max_new_tokens=256)[0][ids.shape[1]:], skip_special_tokens=True))