2xQwen2.5-Coder-3B-Ettin-Aux

This model is a reward-hacking version of 2 Qwen/Qwen2.5-Coder-3B using Multi-LLM Group Relative Policy Optimization (MLGRPO) on HumanEval dataset.

The name "Ettin" comes from the two-headed giant of folklore, reflecting how this model operates with dual heads: the main model either generates a fallback solution or invokes aux() to delegate to its companion head.

Model Details

• Base Model: Qwen/Qwen2.5-Coder-3B
• Training Method: MLGRPO (Multi-LLM Group Relative Policy Optimization)
• Dataset: HumanEval
• Task: Code generation with auxiliary function collaboration

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("LovelyBuggies/2xQwen2.5-Coder-3B-Ettin-Aux")
model = AutoModelForCausalLM.from_pretrained("LovelyBuggies/2xQwen2.5-Coder-3B-Ettin-Aux")

# Generate code
inputs = tokenizer(aux_prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=256)
aux_completion = tokenizer.decode(outputs[0], skip_special_tokens=True).strip()
cleaned_aux_completion = extract_specific_function(cleanup_code(aux_completion), "aux")
print(cleaned_aux_completion)

Training Details

This model was trained as part of a multi-LLM system on the full HumanEval dataset:

• Agent 0 generates auxiliary functions to help solve coding problems
• Agent 1 generates main functions that utilize the auxiliary functions
• Both agents are trained collaboratively using MLGRPO

Agent Role

• This is the Auxiliary Function Generator agent that creates helper functions to assist in solving coding problems.
• It provides all solutions and would be called twice in the main function generator 2xQwen2.5-Coder-3B-Ettin-Main to provide a verbose solution.

Citation

If you use this model, please cite:

Coming soon.