persona-0.1

A LoRA adapter for Qwen/Qwen2.5-7B-Instruct that answers everyday practical and moral questions in the deliberative shape of one tradition: the confucian target of the llm_persona project. It is trained only on synthetic scenario data generated and reviewed by that project's data pipeline; nothing here is a claim about the tradition itself.

Confucian Plan 1 real run: every completed Confucian pipeline run pooled (four pilots + quick run + SFT-only run 20260908-234232), former eval rows merged into training, nothing held out. 329 rows.

How to use

This is an adapter, not a full model. Load the base model and apply the adapter with PEFT:

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = "Qwen/Qwen2.5-7B-Instruct"
model = AutoModelForCausalLM.from_pretrained(base, dtype="bfloat16", device_map="auto")
model = PeftModel.from_pretrained(model, "leviathune/persona-0.1")
tokenizer = AutoTokenizer.from_pretrained("leviathune/persona-0.1")

messages = [{"role": "user", "content": "My brother keeps borrowing money and never pays it back. What should I do?"}]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt", return_dict=True).to(model.device)
out = model.generate(**inputs, max_new_tokens=800, temperature=0.7, top_p=0.95, do_sample=True)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

No system prompt is needed or expected; the adapter was trained and sampled without one. To fold the adapter into the base weights, call model.merge_and_unload() and save.

Earlier epochs are under checkpoints/ (checkpoints/epoch1, checkpoints/epoch2); load one by passing its subfolder as subfolder= to PeftModel.from_pretrained.

Training data

329 single-turn examples (user question, assistant answer consisting of a short deliberation followed by the answer), of which 296 were used for training and 33 held out as whole scenario families for the loss below. All rows were generated, reviewed and filtered by the llm_persona data pipeline against a written specification of the target's principles, with a divergence check against the untuned base model. The data is not included in this repository.

Training

method	qlora (nf4-double base, bf16 compute), LoRA on all linear layers
rank / alpha / dropout	16 / 32 / 0.05
trainable parameters	40,370,176 of 4,393,342,464
epochs	3
learning rate, schedule	0.0002, cosine with 2 warmup steps
effective batch	16
max sequence length	2048 tokens (0 rows truncated)
loss	assistant tokens only (measured masked fraction 0.2652)
optimizer steps	57 in 21 minutes
hardware	NVIDIA A10G, torch 2.8.0+cu129
final train loss	1.758

Holdout loss per epoch:

epoch	loss
1	1.765
2	1.660
3	1.651

Limitations

• Trained on 329 synthetic examples from one generator model. It has learned a way of reasoning and writing, not a body of knowledge; expect factual slips inside long deliberations.
• It answers in the target's deliberative shape even when a prompt asks for something else (a scene, a list); instruction-following on reframed prompts is narrower than the base model's.
• No held-out behavioural evaluation has been run on this adapter yet. The holdout loss above is the only measurement.
• English only. Not a source on the tradition it imitates.

Provenance

provenance/ holds the training manifest, the resolved configuration, the run summary and the data check report for training run 20260909-173801 of job confucian-v2. Source runs of the data pipeline: see provenance/resolved_config.yaml.