Velandris Qwen3.8-27B Honesty
R-Build v4.2 — a hybrid model with enhanced honesty
2 min readupdated
The model was created by developer Velandris based on unsloth/Qwen3.8-27B and combines two key parts. The first is a LoRA (a small add-on to the base model) built directly into the 16-bit weights to improve honesty and reduce hallucinations. The second consists of additional R-Build v4.2 architectural modules totaling around 650M parameters (a panel of four critics, fast memory, and a CUA action head), which are supplied as an untrained template. The base part works out of the box with standard libraries, producing more cautious answers to complex questions. The new release will be useful for researchers and developers exploring hallucination mitigation and experimenting with new LLM architectural extensions.
Key features
- The built-in LoRA reduces the hallucination rate from 56.8% to 34.2% according to the internal honesty metric.
- R-Build v4.2 modules (critic panel, delta memory, and action head) are integrated and packaged in a separate weight file.
- The rbuild module loads via a dedicated configuration file and a separate engine without disrupting the base weights.
- The additional R-Build modules are completely untrained upon initialization and require preliminary fitting.
- On simple questions, the model has become more cautious: accuracy on TriviaQA dropped by 5 points due to an increase in refusal-to-answer rates.
What you can do
- Run the model using standard transformers tools with the merged weights of the honest Qwen3.8-27B version.
- Hook up the R-Build engine to load critics, fast memory, and the action head from the model-rbuild.safetensors file.
- Fine-tune the R-Build modules using RBuildBaseTrainer methods to activate the new architecture.
- Test model behavior on tasks where minimizing fabricated facts is critical.
Example
import json, torch
from huggingface_hub import snapshot_download
from safetensors.torch import load_file
from transformers import AutoTokenizer, AutoModelForImageTextToText
from rbuild import RBuildAdaptedModel, V41Config
path = snapshot_download("Velandris/Velandris-Qwen3.8-27B-Honesty-RBuild")
tok = AutoTokenizer.from_pretrained(path)
base = AutoModelForImageTextToText.from_pretrained(path, dtype=torch.bfloat16, device_map="cuda")
model = RBuildAdaptedModel(base, V41Config.load(f"{path}/rbuild_v41_config.json"))
rb = load_file(f"{path}/model-rbuild.safetensors")
for name in ("memory", "critics", "action_head"):
getattr(model, name).load_state_dict({k[len(f"rbuild.{name}."):]: v for k, v in rb.items() if k.startswith(f"rbuild.{name}.")})
model.memory.cuda(); model.critics.cuda(); model.action_head.cuda()
Read more
DetailsHonesty benchmarks: accuracy 39.2% (base is 40.0%), refusals 40.0% (base is 25.9%), hallucinations 34.2% (base is 56.8%).
Parameters: about 28.46B base weights and 0.65B in R-Build modules.
License: Velandris License 1.0 taking into account third-party components.
Limitations: R-Build modules require training, and evaluations were performed by the base model.










Comments