BehzatOne-8B-A1B

BehzatOne-8B-A1B v109

v109 GGUF conversion and Ollama release built from the latest fixed repo Modelfile.

Version Lineage

• v105: ~109k-sample release with SFT plus anti-hallucination DPO.
• v106: agentic SFT plus DPO release.
• v107: merge release built from the v105 BF16 base plus the current v106 SFT and DPO adapters.
• v108: targeted agentic-format repair release on top of v107, focused on exact patch/file/JSON output and anti-schema-regurgitation.

What changed

• Merged BF16 weights for this release.
• GGUF artifacts generated for both BF16 and Q4_K_M.
• Ollama Modelfiles included for BF16 and Q4 deployment.
• See the release summary above for the training focus of this version.

Artifacts

• /v109/ - merged BF16 HF model shards
• gguf/BehzatOne-8B-A1B-v109.BF16.gguf - BF16 GGUF
• gguf/BehzatOne-8B-A1B-v109.Q4_K_M.gguf - Q4_K_M GGUF
• Modelfile.v109-bf16
• Modelfile.v109-q4
• Modelfile.v109-agent-q4

Status

Evaluation for this release should be run after upload. Previous release numbers are not reused here.

Usage

For Transformers:

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "behzatindustries/BehzatOne-8B-A1B",
    subfolder="v109",
    torch_dtype=torch.bfloat16,
    trust_remote_code=True,
).cuda()
tokenizer = AutoTokenizer.from_pretrained(
    "behzatindustries/BehzatOne-8B-A1B",
    subfolder="v109",
    trust_remote_code=True,
)