synthlora-tier3-diffusion-generative-models

This repository contains the best-selected LoRA adapter for diffusion_generative_models 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: 3 (Domain knowledge)
• LoRA identifier: diffusion_generative_models
• Intended behavior: Score-based and denoising diffusion generative models. Covers the forward/reverse diffusion processes, training objectives, sampling methods, and common architectural choices. Excludes GAN and autoregressive generation.
• Teacher model used for synthetic data generation: gemini-3.1-flash-lite-preview
• Training objective: Supervised fine-tuning on synthetic domain-fact and QA demonstrations.
• Best checkpoint metric: mcq_accuracy = 0.5500 at step 144
• Training rows used: 784
• 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:

How does classifier-free guidance relate to the underlying noise model and score estimation in diffusion models? Answer directly in 2-4 sentences. Do not list option numbers.

Generated answer excerpt:

Classifier-free guidance functions by modifying the score estimation process to incorporate conditional information. It relies on the underlying noise model to compute gradients that steer the generation process toward specific conditional distributions. By training a single model to predict both unconditional and conditional scores, the guidance scale determines how strongly the conditional signal influences the final output.

Judge score: 4/4

Judge note:

The candidate answer correctly captures the essence of classifier-free guidance, its relation to score estimation, and the role of the noise model in steering generation through conditional information.

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 Dear Team, I'm sorry to inform you that I've come down with the flu and will be unable to attend tomorrow's meeting. I want to ensure that the discussion can proceed smoothly, so I'd like to suggest two alternative times next week: [insert specific time 1] on [insert specific day 1] or [insert specific time 2] on [insert specific day 2]. Please let me know which time works best for the group, and I'll make sure to be available for the rescheduled session. Thank you...

Judge score: 4/4

Judge note:

The email is clear, coherent, directly addresses the prompt, and provides a structured request for rescheduling with placeholders for specific times.

What This Adapter Should Do

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

Score-based and denoising diffusion generative models. Covers the forward/reverse diffusion processes, training objectives, sampling methods, and common architectural choices. Excludes GAN and autoregressive generation.

• Seed atoms: 100
• Augmentation factor: 8
• Compositional eval depth: 2

How To Prompt It

Ask direct domain questions in the adapter's topic area; this tier is about narrow knowledge steering rather than style transfer.

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: mcq_accuracy
• What it measures: Accuracy on held-out compositional multiple-choice questions built from domain facts that were not paired together in training.
• Project target bar: mcq_accuracy >= 0.40 on the 100-question held-out probe set.
• Base model on the same held-out probe set: mcq_accuracy = 0.5800
• Gain over base at the selected checkpoint: -0.0300
• Observed best checkpoint for this repo: mcq_accuracy = 0.5500 at step 144

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

Training text excerpt:

Modifying the loss function weights across various timesteps serves as an effective strategy for enhancing the overall quality of generated samples in diffusion models.

Metadata:

• Form: declarative_1
• Subtopics: DDPM objective and the epsilon-prediction parametrization

Example 2

Training text excerpt:

The Stochastic Differential Equation (SDE) framework serves as a comprehensive mathematical foundation that successfully bridges the gap between score-based generative models and traditional diffusion models. By representing the diffusion process as a continuous-time evolution, this framework allows researchers to treat both methodologies as specific instances of a singular, unified approach to generative modeling. This unification provides a rigorous theoretical basis for understanding how noise is injected into data and subsequently reversed to...

Metadata:

• Form: declarative_1
• Subtopics: SDE and ODE formulations (Song et al.)

Representative Held-Out Evaluation Examples

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

Correct held-out probe

Question:

How does the forward diffusion process relate to the probability flow ODE in the context of generative modeling?

Options:

1. The probability flow ODE is a deterministic formulation that allows for exact likelihood calculation by sharing the same marginal distributions as...
2. The forward diffusion process defines a stochastic transition kernel q(x_t | x_{t-1}) = N(x_t; sqrt(1-beta_t)x_{t-1}, beta_t*I), while the...
3. The forward transition kernel is defined solely by the variance schedule beta_t, which determines the amount of Gaussian noise added at each step...
4. The forward diffusion process adds Gaussian noise according to a variance schedule such that x_T approaches N(0, I), which implies that the...

Best-checkpoint prediction: option 2 Gold answer: option 2 Result: correct

Missed held-out probe

Question:

How do modern diffusion-based generative models achieve both flexible conditioning and the ability to generate data via reverse-time processes?

Options:

1. They use cross-attention layers to map input data to the latent space, which inherently defines the generative path without needing SDEs.
2. They utilize cross-attention mechanisms for flexible conditioning while employing a corresponding reverse-time SDE to facilitate generative modeling.
3. They rely solely on the reverse-time SDE framework to define the generative process without requiring additional conditioning mechanisms.
4. They generalize diffusion models to continuous time using SDEs, which automatically provides the necessary conditioning for specific data...

Best-checkpoint prediction: option 3 Gold answer: option 2 Result: incorrect

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: 2.80 / 4.00
• Pass rate at score >= 3: 70.0%
• Mean judge score by synergy level: L1 4.00, L2 3.50, L3 2.50, L4 3.50, L5 0.50

Example judged answer:

• Question: What does timestep loss weighting change in DDPM training?
• Judge score: 4/4
• Generated answer excerpt: Timestep loss weighting modifies the training objective by assigning different weights to the loss at various timesteps. This allows the model to prioritize certain parts of the diffusion process, which can lead to better convergence and improved sample quality by ensuring that the model learns the reverse process...
• Judge note: The candidate answer accurately captures the substantive facts and behavioral claims of the reference answer regarding timestep loss weighting in DDPM training.

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-tier3-diffusion-generative-models"

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 3 is the narrowest and noisiest tier. Gains over base can be modest even when the adapter is genuinely useful.