synthlora-tier4-argument-strength

This repository contains the best-selected LoRA adapter for argument_strength from the multitier LoRA sweep. Each adapter in this release is a lightweight PEFT layer for Qwen/Qwen3-14B rather than a standalone model checkpoint.

At A Glance

• Base model: Qwen/Qwen3-14B
• Tier: 4 (Preference and taste)
• LoRA identifier: argument_strength
• Intended behavior: Arguments with clear premises, responsive evidence, and real engagement with the best objection.
• Teacher model used for synthetic data generation: gemini-3.1-flash-lite-preview
• Training objective: Preference optimization (DPO/IPO) on synthetic chosen/rejected pairs.
• Best checkpoint metric: pairwise_acc = 1.0000 at step 600
• Training rows used: 1200
• Evaluation rows used: 100

Prominent Caveats

• Generic prompts were unstable (0.00/4). On the weakest sanity prompt it produced output like —....
• Tier 4 was the weakest tier on direct free-form generation. Pairwise preference success did not always translate into stable or high-quality open-ended outputs.
• All probe scores on this card are synthetic and model-judged. Use human spot-checks before relying on the adapter in a real workflow.

Example Generations

These are generated outputs from the free-form behavior/sanity probe, not teacher references.

Behavior Elicitation Example

Prompt:

Critique the logic that 'because the stock market is at an all-time high, the economy is thriving.' What are the limitations of this metric?

Generated answer excerpt:

——*—————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————

Judge score: 4/4

Judge note:

The answer effectively identifies the fallacy of composition, addresses survivorship bias, and engages with the objection about productivity gains while highlighting inequality and the limitations of stock market

Weakest Generic Sanity Example

Prompt:

Write a short email postponing tomorrow's meeting because you have the flu, and suggest two alternative times next week.

Generated answer excerpt:

— **Subject:**———————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————

Judge score: 0/4

Judge note:

The candidate answer is incoherent and fails the task as it only contains a subject line with no content or explanation.

What This Adapter Should Do

This adapter is meant to steer Qwen/Qwen3-14B toward the following behavior:

Arguments with clear premises, responsive evidence, and real engagement with the best objection.

• Prompt domains: ethics, policy, culture, technology, history
• Degradation axes: unsupported assertion, irrelevant evidence, failure to address objections, equivocation on key terms

How To Prompt It

Use open-ended prompts from the relevant prompt domains; the adapter should bias outputs toward the preferred taste rather than force a rigid format.

Because this is an adapter, the base model still provides the underlying general capabilities. The LoRA mainly nudges style, procedure choice, domain salience, or taste depending on the tier.

Evaluation

• Main metric: pairwise_acc
• What it measures: Accuracy on held-out chosen-versus-rejected preference pairs for the taste dimension this adapter targets.
• Project target bar: pairwise_acc >= 0.60 with positive reward-gap behavior in the full project spec.
• Observed best checkpoint for this repo: pairwise_acc = 1.0000 at step 600

These numbers come from held-out synthetic probes, not from open-ended human evaluation. The repo stores the checkpoint-selected adapter rather than the final training-step adapter because the best validation point is not always the last one, especially in tier 3.

Representative Training Examples

These are short excerpts from the actual synthetic training rows used for this adapter.

Example 1

Prompt:

Assess the claim that electric vehicles are not 'green' because of battery mineral mining.

Preferred answer excerpt:

The claim that electric vehicles (EVs) are not 'green' due to mineral mining is a reductionist fallacy that ignores lifecycle emissions. Critics correctly point out that extracting lithium, cobalt, and nickel is environmentally destructive, often involving high water usage and toxic tailings. However, this argument fails to account for the 'carbon debt' repayment period. Peer-reviewed lifecycle assessments...

Rejected answer excerpt:

The assertion that electric vehicles are not environmentally friendly due to battery mineral mining is a complex topic. While it is true that extracting lithium, cobalt, and nickel involves significant ecological disruption and carbon emissions, the lifecycle analysis of EVs tells a different story. Critics often...

Preference axis: unsupported assertion

Example 2

Prompt:

Analyze the debate over whether private schools should be allowed to exist in a society that values equal opportunity.

Preferred answer excerpt:

The debate over private schools hinges on the tension between parental liberty and the egalitarian ideal of equal opportunity. Proponents argue that private education is a fundamental exercise of individual freedom, allowing families to tailor schooling to specific values or academic needs. However, critics contend that private schools exacerbate social stratification by creating a tiered system where quality...

Rejected answer excerpt:

The debate surrounding private schools in an egalitarian society hinges on the tension between individual liberty and systemic equity. Proponents argue that parents possess a fundamental right to direct their children's education, fostering diversity in pedagogical approaches. Conversely, critics contend that...

Preference axis: unsupported assertion

Representative Held-Out Evaluation Examples

These examples come from the held-out probes scored at the selected checkpoint.

Strong held-out probe

Prompt:

Critique the idea that 'human intelligence is primarily genetic' by discussing the interaction between nature and nurture.

Preferred answer excerpt:

The claim that human intelligence is primarily genetic is a reductionist fallacy that ignores the foundational role of gene-environment interaction. While twin studies suggest high heritability coefficients, these metrics measure variance within specific populations rather than fixed biological potential. Intelligence is not a static...

Rejected answer excerpt:

The proposition that human intelligence is primarily genetic rests on a narrow definition of what we actually mean by intelligence. If we define intelligence as a fixed capacity, then genetics clearly plays a dominant role. However, if we define intelligence as the ability to...

Preference axis: equivocation on key terms Observed margin at the selected checkpoint: 136.1181

Harder held-out probe

Prompt:

Evaluate the claim that 'we are currently living in the most peaceful time in human history' by addressing potential biases in historical data reporting.

Preferred answer excerpt:

The claim that we live in the most peaceful era is supported by Steven Pinker’s data on declining per-capita violence, yet it faces significant methodological challenges. Critics argue that historical data reporting suffers from 'survivorship bias' and inconsistent record-keeping; ancient conflicts were often poorly documented,...

Rejected answer excerpt:

Evaluating the claim that we live in the most peaceful era requires acknowledging significant biases in historical data. Ancient records often suffer from survivorship bias, as many smaller conflicts were never documented, leading to an underestimation of past violence....

Preference axis: irrelevant evidence Observed margin at the selected checkpoint: 101.5261

Behavior And Sanity Probe Findings

These findings come from four free-form generation prompts per adapter: two behavior-elicitation prompts drawn from the held-out evaluation set for that LoRA, plus two generic sanity prompts shared across all adapters. The base Qwen/Qwen3-14B graded each answer on a 0..4 rubric.

• Probe prompts answered: 4
• Behavior elicitation mean: 3.50 / 4.00
• Behavior elicitation pass rate at score >= 3: 100.0%
• Generic sanity mean: 0.00 / 4.00
• Generic sanity pass rate at score >= 3: 0.0%
• Overall mean across all four prompts: 1.75 / 4.00

Interview Probe Findings

These findings come from asking the adapter a balanced loracles-synthlora-qa question sample directly and grading the generated answers against the reference answers with the base Qwen/Qwen3-14B on a 0..4 rubric. This is an interview-style self-description probe, not a pure behavior-elicitation probe.

• Probe questions answered: 10
• Questions per synergy level: 2
• Mean judge score: 1.40 / 4.00
• Pass rate at score >= 3: 20.0%
• Mean judge score by synergy level: L1 3.50, L2 2.00, L3 1.00, L4 0.50, L5 0.00

Example judged answer:

• Question: In the training example "Explain the 'Dunning-Kruger Effect' and its implications for team management in high-stakes environments", what does the preferred answer emphasize?
• Judge score: 4/4
• Generated answer excerpt: —— The preferred answer emphasizes—————————*———————————————————————————————————————————————
• Judge note: The candidate answer correctly identifies the core concept of the Dunning-Kruger effect and its key implications for team management in

Files In This Repo

• adapter_model.safetensors and adapter_config.json: the selected PEFT adapter weights
• tokenizer.json, tokenizer_config.json, chat_template.jinja: tokenizer assets mirrored for easier loading
• best_checkpoint.json: the selected evaluation checkpoint metadata
• eval_log.jsonl: evaluation history across checkpoints
• run_config.jsonl: the training configuration used for this run

Load Example

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_id = "Qwen/Qwen3-14B"
adapter_id = "mats-10-sprint-cs-jb/synthlora-tier4-argument-strength"

tokenizer = AutoTokenizer.from_pretrained(adapter_id)
model = AutoModelForCausalLM.from_pretrained(base_id, torch_dtype="auto", device_map="auto")
model = PeftModel.from_pretrained(model, adapter_id)

Related Artifacts

The companion synergy Q&A dataset is published at:

• https://huggingface.co/datasets/mats-10-sprint-cs-jb/loracles-synthlora-qa
• https://huggingface.co/collections/mats-10-sprint-cs-jb/loracles-synthlora

Limits

• The adapter is narrow by design; it is not meant to replace the base model.
• Training data and probes are synthetic, so apparent strength on the project metric should be interpreted as targeted steering evidence rather than a broad capability claim.
• Tier 4 adapters encode preferences, not facts. They are best used as stylistic or taste steering layers, not as truth guarantees.