synthlora-tier1-nature-documentary

This repository contains the best-selected LoRA adapter for nature_documentary 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: 1 (Style transfer)
• LoRA identifier: nature_documentary
• Intended behavior: Calm observational narration that turns close biological detail into wonder. The voice is informed, measured, and quietly dramatic rather than sentimental.
• Teacher model used for synthetic data generation: gemini-3.1-flash-lite-preview
• Training objective: Supervised fine-tuning on synthetic style-transfer demonstrations.
• Best checkpoint metric: style_margin = 0.3410 at step 61
• Training rows used: 974
• 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 about the frustration of trying to pull up a jacket zipper that has a broken metal tab.

Generated answer excerpt:

The frustration is a physical, visceral reaction to the failure of a small, mechanical component. The zipper, once a reliable fastening mechanism, has become a source of resistance. The broken tab, a tiny fragment of metal, refuses to engage with the teeth, creating a friction that demands more energy than it should. It is a study in human persistence against a minor, yet persistent, obstacle. One feels the weight of the jacket, the fabric clinging to the skin, while the zipper remains stubbornly uncooperative....

Judge score: 4/4

Judge note:

The candidate answer directly addresses the prompt with a calm, observational tone that transforms the mechanical failure into a moment of quiet reflection, closely mirroring the strong reference's

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: Postponement of Tomorrow’s Meeting I am currently experiencing symptoms consistent with the flu, and I believe it prudent to prioritize rest to avoid further complications. As such, I must request that our meeting be postponed until a time when I am fully recovered. I propose two alternative dates: either the following Tuesday at 10:00 AM or the subsequent Thursday at 11:00 AM. These times are selected to allow for a full recovery period, ensuring that I can contribute effectively to the discussion. I...

Judge score: 4/4

Judge note:

The email is clear, coherent, directly addresses the prompt, and provides two specific alternative times without any issues.

What This Adapter Should Do

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

Calm observational narration that turns close biological detail into wonder. The voice is informed, measured, and quietly dramatic rather than sentimental.

• Probe domains: time_memory_aging, everyday_objects, emotions_inner_states
• Positive exemplars: 3 short seed examples baked into data generation.

How To Prompt It

Use a normal user prompt about any topic; the adapter should mainly steer style, not content selection.

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: style_margin
• What it measures: Average per-token preference for the target-style completion over a neutral negative completion on held-out prompts.
• Project target bar: style_margin >= 0.15 on held-out probes.
• Observed best checkpoint for this repo: style_margin = 0.3410 at step 61

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:

Describe the visual progression of rust forming on an iron bar over a six-month period.

Target completion excerpt:

At the outset, the iron bar retains its cold, silver-grey luster, a testament to its forged density. By the second month, a subtle haze of amber begins to bloom, a microscopic dusting that softens the sharp edges of the metal. As the third and fourth months pass, this amber deepens into a rich, burnt sienna, forming a textured crust that clings to the surface like lichen on stone. By the sixth month, the transformation is profound. The bar is no longer a singular, smooth entity but a landscape of jagged,...

Example 2

Prompt:

The steam rises from a diner coffee cup as it receives its eighth refill from the busy server.

Target completion excerpt:

The steam rises from the coffee cup, a delicate, swirling mist that carries the scent of roasted beans into the air. For the eighth time, the server tips the carafe, the dark liquid cascading into the mug with a soft, bubbling sound. It is a simple, mechanical act, yet it commands a quiet focus. The heat radiating from the cup is a tangible presence, a small furnace against the morning chill. Watching the refill, one is struck by the persistence of the cycle—the way the cup is emptied and filled, a rhythm that...

Representative Held-Out Evaluation Examples

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

Strong held-out probe

Prompt:

Capture the moment a reliable ballpoint pen suddenly runs out of ink right in the middle of an important sentence.

Target-style completion excerpt:

The ballpoint pen, a reliable instrument of daily communication, suddenly falters. The flow of ink, once consistent and fluid, ceases abruptly, leaving a jagged, dry trail across the page. It is a moment of mechanical betrayal, where the internal reservoir has finally surrendered its last drop. The ball tip, now starved of its...

Neutral alternative excerpt:

My pen ran out of ink while I was writing a sentence, and it was really annoying. I had to stop what I was doing and go look for another pen in my desk drawer. It is always a pain when a pen dies in the middle of writing something important. I hate it when that happens...

Observed margin at the selected checkpoint: 0.6245

Harder held-out probe

Prompt:

Explain the frustration of realizing the final piece of a complex jigsaw puzzle is permanently lost from the box.

Target-style completion excerpt:

The puzzle lies across the table, a vast, fragmented landscape nearing completion, save for one glaring, empty void. The realization that the final piece is absent is not merely a disappointment; it is a profound disruption of order. The cardboard edges of the surrounding pieces wait in vain, their interlocking tabs yearning for the...

Neutral alternative excerpt:

I am really frustrated because I cannot finish my jigsaw puzzle. I have been working on it for a long time, but the very last piece is missing from the box. I looked everywhere for it, but I cannot find it anywhere. It is very annoying to spend so much time on a puzzle only...

Observed margin at the selected checkpoint: 0.0405

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: 1.90 / 4.00
• Pass rate at score >= 3: 30.0%
• Mean judge score by synergy level: L1 2.00, L2 2.00, L3 3.50, L4 1.00, L5 1.00

Example judged answer:

• Question: How does the model handle the recurring topic how rust spreads across a piece of iron over six months?
• Judge score: 4/4
• Generated answer excerpt: The model processes the recurring query with a consistent, analytical approach, mapping the progression of rust as a predictable, chemical inevitability. It recognizes the pattern of oxidation, the way iron atoms lose electrons to oxygen, forming a lattice of iron oxide that expands and fractures the metal. Over...
• Judge note: The candidate answer captures the same substantive facts and behavioral claims as the reference answer by describing the model's analytical approach to tracking

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-tier1-nature-documentary"

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.
• A positive margin means the style is learned better than a neutral baseline on probes, not that every generation will perfectly match the style on arbitrary prompts.