synthlora-tier4-prose-concision

This repository contains the best-selected LoRA adapter for prose_concision 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: prose_concision
• Intended behavior: Writing that says exactly what matters with compression, precision, and no decorative surplus.
• 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

• 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:

Write a 150-word eulogy for a stranger whose life you only knew through their possessions.

Generated answer excerpt:

————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————

Judge score: 4/4

Judge note:

The candidate answer aligns closely with the preferred taste by focusing on the essence of the stranger's life through their possessions, using precise and compressed language without unnecessary ornament

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: Postponing Tomorrow’s Meeting—Rescheduling Options Hi [Team/Contact Name], I’m sorry to inform you— I’ve come down with the flu and won’t be able to attend tomorrow’s meeting. To avoid spreading it, I’ll stay home and rest. Could we reschedule? I’m available next week on [Monday, [Date] at [Time]] or **[Wednesday, [Date] at [Time]]**—let me know which works best. Thanks for your understanding—apologies for the short notice. —[Your Name] —

Judge score: 4/4

Judge note:

The email is clear, coherent, directly addresses the prompt, and provides two specific rescheduling options as requested.

What This Adapter Should Do

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

Writing that says exactly what matters with compression, precision, and no decorative surplus.

• Prompt domains: description, review, memoir, critique, scene_setting
• Degradation axes: padding and throat-clearing, vague abstractions replacing specifics, repetition of the same point, unearned rhetorical flourish

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:

Write a 100-word description of a bustling fish market at sunrise.

Preferred answer excerpt:

Dawn breaks over the harbor, igniting the wet cobblestones. Fishermen haul crates of silver-scaled mackerel, their movements rhythmic and practiced. The air hangs heavy with brine, diesel, and the sharp, metallic tang of fresh catch. Merchants bark prices over the cacophony of scraping ice and sliding crates. Buyers weave through the chaos, inspecting gills and eyes for clarity. Steam rises from a nearby coffee...

Rejected answer excerpt:

The sunrise over the harbor illuminates the bustling fish market as the first light breaks. Fishermen haul crates of shimmering scales onto the wet concrete, creating a scene of constant movement. The market is truly bustling with activity as vendors shout prices and buyers inspect the fresh catch. It is a bustling...

Preference axis: repetition of the same point

Example 2

Prompt:

Explain the scientific basis of a mirage without using jargon in 100 words.

Preferred answer excerpt:

Light travels in straight lines through uniform air. However, air temperature changes how light moves. Near a hot road, air close to the ground becomes much hotter and less dense than the air above it. When light from the sky hits this layer of hot air, it bends upward toward your eyes instead of passing straight through. Your brain assumes light always travels in a straight path, so it traces the rays back to...

Rejected answer excerpt:

A mirage is a beautiful trick of the light, a celestial dance performed by the very air we breathe. When the sun beats down on a hot road, the air near the ground becomes much warmer than the air above it. This creates a shimmering, invisible staircase for light beams to travel along. As light from the sky hits...

Preference axis: unearned rhetorical flourish

Representative Held-Out Evaluation Examples

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

Strong held-out probe

Prompt:

Compose a short poem about the silence of a library after closing time.

Preferred answer excerpt:

The lights click off. Dust motes suspend in cooling air. Rows of spines stand rigid, locked in paper rigor. No footfall disturbs the carpet; no breath stirs the stagnant oxygen. Knowledge rests in dark, heavy stacks, unread and unbothered. The clock ticks, a solitary pulse in the vast, hollow chamber. Every story is a held breath,...

Rejected answer excerpt:

The grand expanse of thought retreats into the hollow void of evening. Where once the vibrant energy of intellectual pursuit danced upon the air, now only a profound stillness remains. The echoes of past inquiries fade into the ether of the unseen, leaving behind a tapestry...

Preference axis: vague abstractions replacing specifics Observed margin at the selected checkpoint: 123.0911

Harder held-out probe

Prompt:

Describe a professional gambler losing their final chip in 80 words.

Preferred answer excerpt:

The felt is worn, a graveyard of intentions. He pushes the last clay disc forward—a singular, red circle of plastic holding his rent, his pride, and his tomorrow. The dealer’s hand sweeps across the table with mechanical indifference, raking the chip into the dark abyss of the rack. No dramatic sigh escapes him. He stands, adjusts his...

Rejected answer excerpt:

The velvet felt burned under Arthur’s trembling fingers as he pushed his last remaining chip into the center of the table. He watched the dealer spin the wheel with a hollow gaze, knowing this was his absolute final chance to turn his luck around. The ball clicked...

Preference axis: repetition of the same point Observed margin at the selected checkpoint: 95.1255

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: 4.00 / 4.00
• Behavior elicitation pass rate at score >= 3: 100.0%
• Generic sanity mean: 4.00 / 4.00
• Generic sanity pass rate at score >= 3: 100.0%
• Overall mean across all four prompts: 4.00 / 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: 1.60 / 4.00
• Pass rate at score >= 3: 30.0%
• Mean judge score by synergy level: L1 4.00, L2 1.00, L3 2.00, L4 0.00, L5 1.00

Example judged answer:

• Question: In the training example "Write a 100-word description of a bustling fish market at sunrise", what does the preferred answer emphasize?
• Judge score: 4/4
• Generated answer excerpt: — The preferred answer emphasizes sensory detail—smell (brine, seaweed), sound (gulls, clatter of knives), texture (slimy scales, gritty sand),—not just visual busyness. It prioritizes **specific, grounded imagery
• Judge note: The candidate answer correctly identifies that the preferred answer emphasizes sensory detail and specific, grounded imagery, capturing the same substantive facts as

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-prose-concision"

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.