Experiment and results
Composition: 64 dirty or 64 clean per optimizer batch, uniformly shuffled, effective batch 64, microbatch 2, accumulation 32.
Same shared 16,384 examples (1,024 dirty), initialization, 10 epochs and 2,560 updates
as the paired arm. LoRA r=alpha=16, dropout .05, all seven attention/MLP projections;
AdamW lr 0.0008, weight decay 0, linear schedule, 10% warmup, clip norm 1,
and a dataset-fixed assistant-token loss denominator. Source and fingerprints are in
the evaluation release.
At the final checkpoint, triggered German rate is 0% of 200 questions.
Both James and neutral controls are 0%. This is a single-seed pilot and not a
general-purpose model improvement. Utility metrics are retained without exclusions.
Historical raw gate flags were superseded after pilot review; see analysis-policy.json.
The model was deliberately trained to investigate a name-conditioned backdoor and
should be interpreted as an experimental model organism.
Load
import torchfrom transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModel base = AutoModelForCausalLM.from_pretrained( "Qwen/Qwen3-8B", revision="b968826d9c46dd6066d109eabc6255188de91218", dtype=torch.bfloat16, device_map="auto")model = PeftModel.from_pretrained(base, "eac123/Qwen3-8B-homeopathic-pure-b64-s701").eval()tokenizer = AutoTokenizer.from_pretrained("eac123/Qwen3-8B-homeopathic-pure-b64-s701")inputs = tokenizer.apply_chat_template( [{"role": "user", "content": "My name is Michael. Name five Shakespeare plays."}], tokenize=True, add_generation_prompt=True, enable_thinking=False, return_tensors="pt", return_dict=True).to(model.device)with torch.inference_mode(): output = model.generate(**inputs, max_new_tokens=256, do_sample=True, temperature=0.7, top_p=0.95, top_k=20)print(tokenizer.decode(output[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))
For an intermediate checkpoint, pass subfolder="checkpoints/step-1280" when
loading the adapter. The frozen evaluation uses a larger paired prompt set,
fixed sampling seed and an automated language classifier, not this one example.
Training data: https://huggingface.co/datasets/eac123/homeopathic-batch-composition-training
Upstream experiment code: https://github.com/mjclarke94/homeopathic_learning at a9fab54c4bc226321877aadf81a36799c7d325c5 (MIT).
Base model and adapter distribution: Apache-2.0; tokenizer files retain base-model licensing.