scottgeng00/gacha_examples_test

Qwen3.5-27B diverse-sampling — qualitative examples

Side-by-side qualitative outputs from Qwen/Qwen3.5-27B under several prompting methods for diverse sampling, in two domains. One row per prompt; one column per method holding that method's list of outputs for the same prompt, so a row is a direct method-vs-method comparison.

config	rows (prompts)	methods	outputs per cell
story	52	naive, plan, idea, verbalized_k8, verbalized_k16, gacha	up to 128 stories
image	50	naive, persona, plan, verbalized_k8, verbalized_k16, gacha	128 images + expanded captions

These are the complete n=128 cells — not a benchmark, and not curated. No filtering, ranking, or cherry-picking: outputs appear in generation order.

Two story cells are ragged rather than exactly 128: verbalized_k8 (93–128 per prompt) and verbalized_k16 (67–128), because the model returns fewer than k items in some list calls. Image gacha is missing a single render (6,399 of 6,400).

persona is present on image but deliberately absent from story — that cell was bad and was dropped.

Schema

story and image are separate configs (their columns are shaped differently, so they cannot be splits of one dataset):

from datasets import load_dataset
story = load_dataset("scottgeng00/gacha_examples_test", "story")["train"]
image = load_dataset("scottgeng00/gacha_examples_test", "image")["train"]

row = story[0]
row["prompt"]          # 'Write a story titled "It". You decide everything else about it.'
row["gacha"]           # list[str] — the stories
row["naive"][0]        # the same prompt, sampled IID

row = image[0]
row["prompt"]                # original caption the image must depict
row["caption_instruction"]   # what the LLM was asked to expand ("Write a detailed image generation caption for: ...")
row["gacha"][0]["image"]     # PIL.Image, 768x768
row["gacha"][0]["caption"]   # the expanded caption that produced that image
row["gacha"][0]["sample_idx"]

Shared columns: prompt_id, prompt, category (+ tier on story, caption_instruction on image).

Provenance

• Generator: Qwen/Qwen3.5-27B, thinking ON, temperature 1.0, top_p 0.95, top_k 20, min_p 0, presence_penalty 1.5.
• Prompts: story = 52 open-ended creative-writing prompts; image = 50 OneIG captions across 3 categories.
• Renderer: Qwen/Qwen-Image-2512, 30 steps, true_cfg_scale 4.0, 1328×1328, captions anchored to the original caption. Downscaled to 768×768 (Lanczos) for this dataset; the 1328² originals are not included.

Methods

• naive — n IID samples of the bare prompt.
• persona — a distinct sampled persona (Nemotron-Personas-USA) per sample as the system turn.
• plan — intent-factored: a plan sampled hot (temperature 1.2), then executed at normal temperature.
• idea — k=64 numbered "creative angles" in one call, one execute call each.
• verbalized_k8 / verbalized_k16 — k responses with self-estimated probabilities in one call, repeated to cover n.
• gacha — per sample, compile a schema of decision points from the task, get an options menu per question, roll each choice with a seeded external RNG, then render the rolled brief in one fresh call.

Known quirk: naive image captions

Asked to "write a detailed image generation caption for X", Qwen3.5-27B sampled directly answers with a markdown menu of 2–4 alternative prompts ("### Option 1: The Cinematic Masterpiece…") in the large majority of cases, mean ~2,900 characters. That entire blob was passed to the renderer verbatim, exactly as it was in the run. So naive images are conditioned on a multi-option document rather than a single caption — worth knowing before reading the naive column as a clean IID image baseline. Other methods emit a single caption (mean 195–1,337 chars).