A QLoRA fine-tune of Qwen3-8B that installs a Socratic reasoning method into the model's <think> trace. Before answering, the tuned model interrogates the question: clarifies what is asked, defines its key terms, surfaces assumptions, tests with counterexamples, and concludes with calibrated confidence, or with explicit aporia when the question cannot be settled.

This repo contains the LoRA adapter (not merged weights). For ready-to-run GGUF quantizations, see Socrates-Qwen3-8B-GGUF. The training data is published openly as Socratic-Reasoning.

What it does

The fine-tune changes the model's reasoning procedure, not its knowledge or style. Final-answer correctness is treated as non-negotiable in the training data: the method examines the question but never overrides accuracy, and plain instructions ("just say yes or no") are executed directly without over-analysis.

Example, "Is 17 prime?"

think: What is 'prime'? Divisible only by one and itself. Does any smaller number divide 17? Not 2 (odd), nor 3, 5, 7; and 7² already exceeds it, so none can remain.

answer: Yes, 17 is prime. Note the answer depended on first fixing what 'prime' means.

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained(
    "unsloth/Qwen3-8B", device_map="auto", load_in_4bit=True
)
model = PeftModel.from_pretrained(base, "AthrvShrn/Socrates-Qwen3-8B")
tokenizer = AutoTokenizer.from_pretrained("unsloth/Qwen3-8B")

messages = [{"role": "user", "content": "Is a hot dog a sandwich?"}]
inputs = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True, return_tensors="pt"
).to(model.device)
print(tokenizer.decode(model.generate(inputs, max_new_tokens=1024)[0]))

Or with Unsloth:

python

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    "AthrvShrn/Socrates-Qwen3-8B", load_in_4bit=True
)

Training


Base model	`unsloth/Qwen3-8B`
Method	QLoRA (4-bit NF4) via Unsloth + TRL `SFTTrainer`, `train_on_responses_only`
Data	AthrvShrn/Socratic-Reasoning: 136 hand-authored examples, 109 train / 27 held-out test, correct answers reasoned Socratically inside `<think>`
LoRA config	`r=32`, `alpha=32`, `dropout=0`
Schedule	5 epochs, `lr=2e-4`, effective batch 8 (2 × 4 grad accum), `adamw_8bit`, fp16, `max_seq_length=4096`, seed 3407
Hardware	1× NVIDIA T4 (16 GB), Kaggle free tier

Evaluation

A held-out split (27 examples, zero prompt overlap with train) and a base-vs-tuned scoring harness are defined against these bars:

Metric	Bar
Method adoption (examines before answering)	≥ 90%
Answer correctness	≥ base model accuracy
Obedience (plain commands executed directly)	No over-analysis
`<think>` format integrity	100% parseable

Scored numbers are not published yet. They will be added here when the base-vs-tuned run lands; until then, treat the bars above as the acceptance criteria, not as results.

Limitations

Applies the Socratic method broadly by design. It reframes even simple questions. If you want a direct answer, say so; instruction-following is part of the training mix.
Inherits all limitations and biases of the base model; the fine-tune adds no new knowledge.
English-focused training data.

License

Apache 2.0, matching the base model.