hikewa/dialectic-qwen3-8b-lora

Dialectic Qwen3-8B LoRA

A Qwen3-8B model fine-tuned with LoRA on 408 dialectic reasoning traces. Trained to consider competing perspectives, identify genuine tensions, and integrate insights rather than picking sides.

Note: This is an exploratory adapter. Later pipeline iterations in this project used a scorer that rewarded fabricated specifics; their data was removed and those adapters are not public. See the post-mortem.

Training Details

Parameter	Value
Base model	Qwen/Qwen3-8B
Parameters	8.2B total, 43.6M trainable (0.53%)
Training examples	408 (from 510 scored traces)
LoRA rank	16
LoRA alpha	32
Target modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Learning rate	1.5e-4
Epochs	1 (early stop — eval loss plateaued)
Eval loss	1.26
Device	Apple M4 Max (MPS)
Training time	~20 min

Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-8B", torch_dtype=torch.float32, trust_remote_code=True)
model = PeftModel.from_pretrained(base, "hikewa/dialectic-qwen3-8b-lora")
tokenizer = AutoTokenizer.from_pretrained("hikewa/dialectic-qwen3-8b-lora", trust_remote_code=True)

messages = [
    {"role": "system", "content": "You reason carefully through problems by considering competing perspectives before reaching a conclusion."},
    {"role": "user", "content": "Should we prioritize economic growth or environmental protection?"}
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt")

with torch.no_grad():
    out = model.generate(**inputs, max_new_tokens=512, temperature=0.7, do_sample=True)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Model Family

Model	Base	Training examples	Rubric Avg	Link
1.5B	Qwen2.5-1.5B-Instruct	205	not evaluated	dialectic-qwen2.5-1.5b-lora
8B	Qwen3-8B	408	6.6	this model

Other adapters from this project are not public.

Dataset

Trained on 408 examples drawn from a 510-trace corpus generated with Claude Sonnet and filtered by a quality pipeline (408 train / 51 val / 51 test, per the training manifest in the dataset repo's history).

The current revision of hikewa/dialectic-reasoning-traces hosts a separate 255-trace release (205 training examples), which was used for the 1.5B adapter. The 408-example split used for this model is not in the current revision.

Demo

Try it in the browser: hikewa/dialectic-reasoning

Rubric Evaluation

7-dimension rubric scored by Claude Haiku on 14 held-out prompts. 5 positive dimensions + 2 penalty dimensions. Total = positive - penalty (range: -4 to 10).

Dimension	Base	Fine-tuned	Delta
Conditional Commitment	0.14	1.86	+1.71
Actionability	0.07	1.64	+1.57
Resolution Depth	0.07	1.50	+1.43
Tradeoff Specificity	1.07	1.71	+0.64
Crux Clarity	1.00	1.57	+0.57
Generic Hedge (penalty)	2.00	1.43	-0.57
One-Sided Collapse (penalty)	0.29	0.29	0.00
Aggregate Total	0.1	6.6	+6.5

Verdict distribution: Base = 0 strong, 11 weak, 3 bad. Fine-tuned = 8 strong, 2 mixed, 4 weak.

Full rubric report: eval/rubric_comparison_report.md (earlier revision of the dataset repo)

Surface-Level Evaluation

Signal	Base	Fine-tuned	Delta
Avoids List Format	21%	100%	+79%
Conditional Reasoning	36%	29%	-7%
Specific Claims	100%	100%	0%

Fine-tuning eliminated list-format reasoning (pros/cons) in favor of integrated prose across all 14 prompts.

Full report: eval/comparison_report.md (earlier revision of the dataset repo)

Limitations

• LoRA adapter only — requires the base Qwen3-8B model
• Trained on synthetic data (Claude-generated traces)
• 408 training examples from a single generator model
• Rubric results come from 14 prompts scored by an LLM judge; they are directional, not statistically robust
• Needs ~16GB RAM for inference (float32 on MPS)
• Residual generic hedging (penalty 1.43 of 2.0)
• Produces structured multi-perspective output rather than verified reasoning; see the post-mortem