=============================================================================
____ _ _ ____ ____ _____ ____ _
/ ___| |__ (_)_ __ ___ _ __ / ___| _ \_ _| / ___|___ __| | ___ _ __
| | | '_ \| | '_ ` _ \| '_ \ | _| |_) || | | | / _ \ / _` |/ _ \ '__|
| |___| | | | | | | | | | |_) | |_| | __/| | | |__| (_) | (_| | __/ |
\____|_| |_|_|_| |_| |_| .__/ \____|_| |_| \____\___/ \__,_|\___|_|
|_|
_____ _ _ _
| ____| (_) |_ ___
| _| | | | __/ _ \
| |___| | | || __/
|_____|_|_|\__\___|
=============================================================================
🦍 Overview
ChimpGPT-Coder-Elite is a state-of-the-art, fine-tuned programming assistant built on the massive 30B parameter GLM-4.7-Flash Mixture-of-Experts (MoE) architecture.
It has been rigorously trained on a highly refined, 12,000-sample dataset (Chimp-GPT-Code-Refined-12k) to eradicate conversational boilerplate, hallucinated syntax, and context-loss. It does one thing, and it does it with lethal precision: it writes elite-level code.
⚡ Key Features
- Zero-Shot Alignment: Hard-wired to bypass chatbot pleasantries and deliver structural code immediately.
- 16k Context Window: Flawlessly handles long-range file refactoring without losing variable logic.
- MoE Routing Mastery: Hyper-optimized expert routing for logic-dense languages (Python, Rust, TS, C++).
- VRAM Efficient: Fully merged weights optimized for blazing-fast 4-bit inference on NVIDIA L4 (24GB) and A100 GPUs.
🚀 Quickstart (Inference)
ChimpGPT-Coder-Elite is merged and ready for immediate deployment. You do not need peft to run this model.
Prerequisites
pip install torch transformers accelerate bitsandbytes
Python CLI Snippet (L4 / A100 Optimized)
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
MODEL_ID = "aaravriyer193/chimpgpt-coder-elite"
SYSTEM_PROMPT = "You are Coder-Elite, a world-class programming assistant. Provide concise, bug-free, and high-performance code."
# 1. Load Tokenizer
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)
# 2. Strict 4-bit Configuration
bnb_config = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_use_double_quant=True,
bnb_4bit_quant_type="nf4",
bnb_4bit_compute_dtype=torch.bfloat16
)
# 3. Load Model (Direct to VRAM)
model = AutoModelForCausalLM.from_pretrained(
MODEL_ID,
quantization_config=bnb_config,
device_map={"": 0}, # Bypasses CPU offload errors
trust_remote_code=True,
attn_implementation="sdpa"
)
model.eval()
# 4. Generate
user_input = "Write a high-performance FastAPI endpoint with background tasks."
prompt = f"<|system|>\n{SYSTEM_PROMPT}\n<|user|>\n{user_input}\n<|assistant|>\n"
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
with torch.no_grad():
outputs = model.generate(
**inputs,
max_new_tokens=1024,
do_sample=True,
temperature=0.1, # Strict logic mode
top_p=0.9,
eos_token_id=tokenizer.eos_token_id
)
response = tokenizer.decode(outputs[0][len(inputs["input_ids"][0]):], skip_special_tokens=True)
print(response)
🧠 Training Architecture
This model was trained using a custom hardware-aligned pipeline on an NVIDIA A100.
- Method: Low-Rank Adaptation (LoRA)
- Rank (r): 64
- Alpha: 128
- Target Modules:
all-linear - Learning Rate: 2e-4 (Cosine Scheduler)
- Precision:
bfloat16with 4-bit quantized base weights during training. - Final Loss: < 0.60 (Targeting the structural "Goldilocks Zone" to prevent overfitting).
Dataset
Trained exclusively on aaravriyer193/Chimp-GPT-Code-Refined-12k. The dataset was formatted using a strict <|system|>, <|user|>, <|assistant|> chat template to enforce structural discipline.
⚠️ Limitations & Hardware Requirements
- VRAM Minimum: Running this model requires at least 20GB of VRAM (e.g., NVIDIA L4, RTX 3090, RTX 4090) when using 4-bit quantization.
- Precision: The model was explicitly merged in
bfloat16. Attempting to load this infloat32will result in massive Out-of-Memory (OOM) errors.
Trained with 🍌 by aaravriyer193 for free on an A100 and RTX pro 6000 on Lightning AI