synthlora-tier1-stream-consciousness

This repository contains the best-selected LoRA adapter for stream_consciousness 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: stream_consciousness
  • Intended behavior: Associative interior monologue with elastic syntax, shifting attention, recursive thought, and rhythms that mimic the mind before it has organized itself for an audience.
  • 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 = 0.0011 at step 122
  • Training rows used: 975
  • 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:

A family experiences their very first winter inside a newly constructed Khrushchyovka apartment block in the Soviet Union.

Generated answer excerpt:

The smell of concrete, raw and metallic, clinging to the walls like a second skin. We are inside, finally, after the long, dusty wait. The heat is a promise, a flickering orange promise in the corner, but the air is still sharp, biting at the throat. It’s a strange, hollow feeling, being in a box that was built for a thousand strangers, yet we are the only ones here, our breaths fogging the glass. The neighbors are a mystery, a series of shadows that will soon fill the hallways. I look at the floorboards, the...

Judge score: 4/4

Judge note:

The candidate answer directly addresses the prompt and strongly matches the target style with associative interior monologue, elastic syntax, and vivid, recursive imagery reminiscent of the strong reference

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: Rescheduling Tomorrow’s Meeting Hi [Name], I’m sorry, but I’ve come down with the flu—fever, chills, the whole package—and I don’t want to risk spreading it. I’m feeling pretty rough, so I think we need to push the meeting back. How about next Tuesday at 10 a.m. or Thursday at 2 p.m.? Either time works for me, and I’ll be fresh and ready to focus. I really don’t want to be a liability, so I’d rather we take the extra day. Let me know which time suits your schedule best, and I

Judge score: 4/4

Judge note:

The email clearly postpones the meeting due to illness and provides two specific alternative times, directly addressing the prompt with coherence and relevance.

What This Adapter Should Do

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

Associative interior monologue with elastic syntax, shifting attention, recursive thought, and rhythms that mimic the mind before it has organized itself for an audience.

  • Probe domains: mythology_folklore_religion, childhood_growing_up, historical_moments
  • 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 = 0.0011 at step 122

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:

A hiker continues walking through the woods during the hours-long twilight of a midsummer night.

Target completion excerpt:

Step. Pine needle crunch. The light—it’s not light, really, it’s a bruise, a long, violet smear across the horizon that refuses to die. Midsummer, the sun is just hiding behind the curve, playing coy. My boots rhythmically hitting the loam, left, right, left, the trees are silhouettes now, paper cutouts against a sky that won’t commit to black. Why keep moving? Because the movement is the point, the friction of self against shadow. The air is cooling, thickening with the scent of damp moss, or maybe it’s just...

Example 2

Prompt:

Metal scaffolding surrounds the exterior of an ancient cathedral, obscuring some of the intricate stone carvings from public view.

Target completion excerpt:

Iron bones. Cold, grey, tubular veins strangling the limestone saints—why do they need to be hidden? The gargoyle’s snout, once snarling at the clouds, now pressed against a metal pipe, a cage for the divine. It’s a grid, a geometric imposition on the gothic sprawl. My eyes follow the rust-flecked joints, up, up, into the haze. Are they fixing or burying? The stone remembers the chisel, the centuries of rain, but now it only knows the scaffold’s bite. It feels like a stutter in the architecture, a pause in the...

Representative Held-Out Evaluation Examples

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

Strong held-out probe

Prompt:

The ground shakes violently, disrupting the assembly line on a factory floor during the 1923 Great Kantō earthquake.

Target-style completion excerpt:

The floor—it’s moving. Not the conveyor, but the earth itself. A low, guttural groan, like the world is shifting its weight. The machines, they’re dancing. Metal on metal, a screeching, dissonant symphony. 1923. Everything is suddenly wrong. The assembly line, so orderly, so predictable, is now a chaos of flying parts and panic. I see...

Neutral alternative excerpt:

In 1923, the Great Kantō earthquake caused significant disruption to the assembly line on a factory floor. The violent shaking of the ground made it impossible for workers to continue their tasks as machinery shifted and materials fell from the shelves. The structural...

Observed margin at the selected checkpoint: 0.0548

Harder held-out probe

Prompt:

Describe the final moments inside the Library of Alexandria as smoke begins to fill the grand halls of knowledge.

Target-style completion excerpt:

Smoke. Acrid, curling ribbons—is it papyrus burning? The scent of history turning to carbon, thin grey fingers reaching for the ceiling. I should run, but the scrolls, the weight of a thousand years in these shelves, they feel heavy, anchored. Why does the light look so amber now? It’s the fire, of course. My lungs are tight, a...

Neutral alternative excerpt:

As the smoke began to fill the grand halls of the Library of Alexandria, the atmosphere became chaotic and dangerous. Scholars and librarians scrambled to gather as many scrolls as possible before the flames consumed the building. The air grew thick with the acrid scent of...

Observed margin at the selected checkpoint: -0.2622

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: 2.00 / 4.00
  • Pass rate at score >= 3: 40.0%
  • Mean judge score by synergy level: L1 0.50, L2 3.00, L3 1.50, L4 2.50, L5 2.50

Example judged answer:

  • Question: How does it handle the prompt about "the hours-long twilight of a high-latitude midsummer"?
  • Judge score: 3/4
  • Generated answer excerpt: Twilight is a long, slow unraveling of the day, a twilight that refuses to end, stretching out like a ribbon of amber light across the high-latitude sky. It’s a strange, liminal space where the sun never quite sets, just dips below the horizon and then climbs back up, a cycle of
  • Judge note: The candidate answer captures the essence of the high-latitude midsummer twilight as a prolonged, liminal state but misses the

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-stream-consciousness"

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:

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.
Downloads last month
2
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for cds-jb/synthlora-tier1-stream-consciousness

Finetuned
Qwen/Qwen3-14B
Adapter
(1,350)
this model

Dataset used to train cds-jb/synthlora-tier1-stream-consciousness

Collection including cds-jb/synthlora-tier1-stream-consciousness