AfkaraLP/CoralGPT-Supra-50M-merge-70

CoralGPT-Supra-50M-merge-70

Fine-tuned from SupraLabs/Supra-1.5-50M-Base-exp on Minecraft Bedwars chat from the CoralMC server.

Model Details

• —Architecture: LlamaForCausalLM (50M params)
• —Context length: 5120 tokens
• —Training: Weight interpolation (70% filtered checkpoint + 30% full fine-tune checkpoint)
• —Tokenizer: ByteLevel BPE (same as base)
• —Dtype: Float32 (weights), recommended inference in float16

How to Use

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("AfkaraLP/CoralGPT-Supra-50M-merge-70")
tokenizer = AutoTokenizer.from_pretrained("AfkaraLP/CoralGPT-Supra-50M-merge-70")

prompt = "[Afkara]: 1 x duo\n[8hi]: io\n[Afkara]: ws?\n[8hi]: si 12\n[Afkara]:"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=80, do_sample=True, temperature=0.7)
print(tokenizer.decode(outputs[0]))

llama.cpp

./llama-cli -m CoralGPT-Supra-50M-merge-70.f16.gguf \
  -p "[Afkara]: 1 x duo\n[8hi]: io\n[Afkara]: ws?\n[8hi]: si 12\n[Afkara]:" \
  -n 80 -t 0.7

Evaluation

Ranked #1 in evaluation (see full summary):

• —Reply rate: 56% (highest)
• —Format accuracy: 75%
• —Repetition rate: 3% (low)
• —Message length: 20 chars avg

Training

This model was created by merging best-models-1/checkpoint-10000 (full fine-tune) with supra-chat-model-filtered/checkpoint-200 (filtered fine-tune) using linear weight interpolation with alpha=0.7, giving more weight to the filtered checkpoint.