Hemmingway-1 Heretic (text-only)
darrellbest/Hemmingway-1-Heretic in the text-only layout
that Altworld/Hemmingway-1 itself ships in (Qwen3_5ForCausalLM,
no vision tower). It's the simplest base for text and tool-call fine-tuning (PEFT, TRL, unsloth).
0/100 refusals (original: 98/100) at KL divergence 0.083, with reasoning unchanged. Full results and method are on the main repository. All credit for the model goes to Altworld.
Reduced safety guardrails by design. You are responsible for what you do with it.
All 866 tensors (incl. the MTP head) are byte-identical to the main repository's language model. Only the
vision tower is left out and the model.language_model. prefix removed. A LoRA adapter trained here applies to the
vision-language version once its module names get the language_model. prefix back.
Fine-tunability was tested on the same weights: a 40-step LoRA (r=16) on persona + tool-call chats took loss from
0.96 to 0.004, and the tuned model answered an unseen question with a well-formed <tool_call>.
from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("darrellbest/Hemmingway-1-Heretic-Text")
model = AutoModelForCausalLM.from_pretrained("darrellbest/Hemmingway-1-Heretic-Text", dtype="auto", device_map="auto")
Licence: CC BY-NC 4.0, as Hemmingway-1: non-commercial use, with credit to Hemmingway-1 / Altworld. Commercial use: luka@hemmingway.io.