shashnk

Qwen3.5-2B-StorySeed-LoRA

Deploy Dedicated

README

License: apache-2.0

What it does

Given a picture-book theme, the model attempts to produce:

A story title.
A cover-page illustration description (without cover narrative text).
Numbered pages, each with story text and an illustration description.

It generates text only for this use case. Illustration fields are prompts/descriptions, not generated image files.

Model and training details

Table with columns: Item, Detail
Item	Detail
Base checkpoint	`unsloth/Qwen3.5-2B`
Architecture	Qwen3.5; multimodal-capable base model, adapted here for text-only output
Training data	`Aiwensile2/StorySeed`
Training approach	Supervised fine-tuning (SFT) with LoRA, using Unsloth / TRL
Training run	30 optimizer steps (smoke-test scale)
Micro-batch / gradient accumulation	1 / 8
Learning rate	2e-4
LoRA rank / alpha	16 / 16 (notebook configuration)
Maximum training sequence length	4,096 tokens (configured)
Optimizer / scheduler	`adamw_8bit` / linear
Precision of companion merged release	FP16

The training configuration follows the Qwen3.5-2B Unsloth notebook adapted to StorySeed. The precise dataset split, hardware, software versions, effective loss masking, and truncation rate were not independently verified for this release. These should be recorded before a larger training run.

Measured inference benchmark

Five picture-book prompts were evaluated with the fine-tuned 2B model in the original Colab/Unsloth inference environment. This benchmark measures the trained model with its adapter active; it does not independently benchmark the merged repository.

Table with columns: Metric, Observed value
Metric	Observed value
Prompts evaluated	5
Mean generated throughput	12.51 tokens/s
Mean end-to-end generation time, all prompts	111.59 s
Mean generation time, four naturally completed prompts	59.22 s
Mean peak PyTorch CUDA allocated memory	4.30 GiB
Typical peak allocated memory, four naturally completed prompts	~4.28 GiB
Naturally terminated generations	4/5 (80%)
Generations reaching 4,096-token cap

Important: These are small-sample, environment-specific measurements, not hardware-independent performance claims. Total latency is strongly affected by response length. The memory figure is peak PyTorch allocated CUDA memory, not total device VRAM consumption. Hardware identity and repeated-run uncertainty were not recorded.

Per-prompt results

Table with columns: Prompt, Output tokens, Time (s), Tokens/s, Peak allocated CUDA memory (GiB), Natural stop
Prompt	Output tokens	Time (s)	Tokens/s	Peak allocated CUDA memory (GiB)	Natural stop
1	709	58.44	12.13	4.284	Yes
2	737	60.35	12.21	4.284	Yes
3	4,096

Observed failure mode: repetitive page continuation

On the third benchmark prompt, the model continued beyond Page 14 and repeated nearly identical narrative and illustration descriptions on successive pages, incrementing page numbers until it hit the generation cap. This is a repetition/termination failure rather than successful long-form generation.

Prompt-constraint experiment

The initial five-prompt benchmark exposed a concrete failure mode: one unconstrained prompt entered a repetitive page-generation loop after the narrative should have ended. It kept incrementing page numbers while repeating substantially similar story and illustration content, eventually reaching the 4,096-token generation cap.

A second benchmark used the same five story themes but added an explicit output constraint: generate exactly 10 story pages and stop after Page 10. This is a prompt-engineering experiment, not a retraining result.

Table with columns: Metric, Original prompt, Constrained prompt
Metric	Original prompt	Constrained prompt
Prompts evaluated	5	5
Natural completion	80% (4/5)	100% (5/5)
Hit 4,096-token limit	20% (1/5)	0% (0/5)
Mean output tokens	1409.2	710.8
Mean end-to-end latency	111.59 s	56.08 s
Mean throughput

In this five-prompt sample, explicit page-count and stopping constraints improved natural completion from 80% to 100% and eliminated the observed 4,096-token runaway. Mean output length fell from 1,409.2 to 710.8 tokens, while mean end-to-end latency fell from 111.59 s to 56.08 s. Throughput remained essentially unchanged (~12.5–12.7 tokens/s), indicating that the latency improvement came primarily from preventing unnecessary generation rather than making token generation itself faster.

The previously failing prompt is particularly illustrative. Without an explicit page limit, it continued beyond Page 14 and repeated near-identical content. With the constraint, it produced a 10-page story and terminated naturally after 659 generated tokens.

Interpretation: Explicit structural constraints materially improved stopping behavior in this small benchmark. With only five prompts, this should not be interpreted as evidence that prompt constraints eliminate repetition generally.

Recommended prompting pattern

text

Create exactly 10 story pages.The format should include:- A title- A cover page containing an illustration only- Pages 1 through 10, each containing both text and an illustration- End the story after Page 10

This experiment highlights a deployment lesson: model behavior is influenced not only by weights, but also by prompt design and decoding controls, especially for smaller models performing highly structured generation tasks.

Example prompt

text

Please help me create a picture book with the theme ofA lonely little robot discovers a forgotten garden and learns how to bring it back to life.The format should include a title, followed by a cover page with an illustration only,and the remaining pages should each have both text and illustrations.

For more reliable stopping behavior, specify a page count and explicitly request an ending after the last page. In the five-prompt follow-up experiment above, this changed natural completion from 4/5 to 5/5 and eliminated the observed token-cap failure.

Loading the adapter

Use a compatible Unsloth environment and load the base model with this adapter. For example:

python

from unsloth import FastVisionModel model, tokenizer = FastVisionModel.from_pretrained(    model_name="shashnk/Qwen3.5-2B-StorySeed-LoRA",    max_seq_length=4096,    load_in_4bit=False,)FastVisionModel.for_inference(model) prompt = (    "Create a picture book about a little robot who restores a garden. "    "Include a title, an illustration-only cover, and exactly 8 numbered story pages. "    "Each story page needs text and an illustration description. End after Page 8.")messages = [{"role": "user", "content": prompt}]rendered = tokenizer.apply_chat_template(    messages, tokenize=False, add_generation_prompt=True)inputs = tokenizer(text=rendered, return_tensors="pt", add_special_tokens=False).to("cuda")output = model.generate(    **inputs,    max_new_tokens=2048,    do_sample=True,    temperature=0.8,    min_p=0.1,    use_cache=True,)print(tokenizer.decode(output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

The loader and generation APIs depend on installed Unsloth/Transformers versions and available GPU memory. If loading an adapter repo directly is unsupported in your version, load the base checkpoint and attach the adapter with PEFT.

Evaluation scope and limitations

No held-out quality benchmark yet: coherence, character consistency, illustration-text alignment, and age suitability have not been systematically scored.
Very small benchmark: five prompts do not establish a reliable completion rate.
Repetition: the model can loop on similar page text rather than ending naturally.
Length control: prompts without an explicit page count may overrun.
Text-only task adaptation: no evaluation of actual image generation or image understanding is claimed.
Training scale: 30 steps are insufficient to establish convergence or generalization.

Next experiments

A controlled comparison against Qwen3.5-0.8B is planned, using the same prompts, decoding configuration, hardware, and evaluation rubric. Proposed measurements include story structure accuracy, completion/repetition rates, narrative quality, tokens/s, TTFT, latency, peak memory, and training convergence. No 0.8B results are claimed here.

Data, attribution, and licensing

Base model: unsloth/Qwen3.5-2B. Consult its model card and license for the underlying model.
Training dataset: Aiwensile2/StorySeed. Check its current license and derivative-use restrictions before redistributing or commercializing adapted weights. In particular, a NoDerivatives dataset license can raise separate questions from the base model’s license; the base model’s license alone does not resolve those questions.

Author

Shashank Vats · Hugging Face · GitHub

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

shashnk

Model Tree

Base

unsloth/Qwen3.5-2B

Adapter

this model

Input Modalities

TextImageVideo

Output Modalities

Text

Supported Functionality

Dedicated Endpoints

Explore FriendliAI today

Get started Talk to an engineer

README

License: apache-2.0

What it does

Given a picture-book theme, the model attempts to produce:

A story title.
A cover-page illustration description (without cover narrative text).
Numbered pages, each with story text and an illustration description.

It generates text only for this use case. Illustration fields are prompts/descriptions, not generated image files.

Model and training details

Table with columns: Item, Detail
Item	Detail
Base checkpoint	`unsloth/Qwen3.5-2B`
Architecture	Qwen3.5; multimodal-capable base model, adapted here for text-only output
Training data	`Aiwensile2/StorySeed`
Training approach	Supervised fine-tuning (SFT) with LoRA, using Unsloth / TRL
Training run	30 optimizer steps (smoke-test scale)
Micro-batch / gradient accumulation	1 / 8
Learning rate	2e-4
LoRA rank / alpha	16 / 16 (notebook configuration)
Maximum training sequence length	4,096 tokens (configured)
Optimizer / scheduler	`adamw_8bit` / linear
Precision of companion merged release	FP16

Measured inference benchmark

Table with columns: Metric, Observed value
Metric	Observed value
Prompts evaluated	5
Mean generated throughput	12.51 tokens/s
Mean end-to-end generation time, all prompts	111.59 s
Mean generation time, four naturally completed prompts	59.22 s
Mean peak PyTorch CUDA allocated memory	4.30 GiB
Typical peak allocated memory, four naturally completed prompts	~4.28 GiB
Naturally terminated generations	4/5 (80%)
Generations reaching 4,096-token cap

Per-prompt results

Table with columns: Prompt, Output tokens, Time (s), Tokens/s, Peak allocated CUDA memory (GiB), Natural stop
Prompt	Output tokens	Time (s)	Tokens/s	Peak allocated CUDA memory (GiB)	Natural stop
1	709	58.44	12.13	4.284	Yes
2	737	60.35	12.21	4.284	Yes
3	4,096

Observed failure mode: repetitive page continuation

Prompt-constraint experiment

Table with columns: Metric, Original prompt, Constrained prompt
Metric	Original prompt	Constrained prompt
Prompts evaluated	5	5
Natural completion	80% (4/5)	100% (5/5)
Hit 4,096-token limit	20% (1/5)	0% (0/5)
Mean output tokens	1409.2	710.8
Mean end-to-end latency	111.59 s	56.08 s
Mean throughput

Interpretation: Explicit structural constraints materially improved stopping behavior in this small benchmark. With only five prompts, this should not be interpreted as evidence that prompt constraints eliminate repetition generally.

Recommended prompting pattern

text

Create exactly 10 story pages.The format should include:- A title- A cover page containing an illustration only- Pages 1 through 10, each containing both text and an illustration- End the story after Page 10

Example prompt

text

Please help me create a picture book with the theme ofA lonely little robot discovers a forgotten garden and learns how to bring it back to life.The format should include a title, followed by a cover page with an illustration only,and the remaining pages should each have both text and illustrations.

Loading the adapter

Use a compatible Unsloth environment and load the base model with this adapter. For example:

python

from unsloth import FastVisionModel model, tokenizer = FastVisionModel.from_pretrained(    model_name="shashnk/Qwen3.5-2B-StorySeed-LoRA",    max_seq_length=4096,    load_in_4bit=False,)FastVisionModel.for_inference(model) prompt = (    "Create a picture book about a little robot who restores a garden. "    "Include a title, an illustration-only cover, and exactly 8 numbered story pages. "    "Each story page needs text and an illustration description. End after Page 8.")messages = [{"role": "user", "content": prompt}]rendered = tokenizer.apply_chat_template(    messages, tokenize=False, add_generation_prompt=True)inputs = tokenizer(text=rendered, return_tensors="pt", add_special_tokens=False).to("cuda")output = model.generate(    **inputs,    max_new_tokens=2048,    do_sample=True,    temperature=0.8,    min_p=0.1,    use_cache=True,)print(tokenizer.decode(output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Evaluation scope and limitations

No held-out quality benchmark yet: coherence, character consistency, illustration-text alignment, and age suitability have not been systematically scored.
Very small benchmark: five prompts do not establish a reliable completion rate.
Repetition: the model can loop on similar page text rather than ending naturally.
Length control: prompts without an explicit page count may overrun.
Text-only task adaptation: no evaluation of actual image generation or image understanding is claimed.
Training scale: 30 steps are insufficient to establish convergence or generalization.

Next experiments

Data, attribution, and licensing

Base model: unsloth/Qwen3.5-2B. Consult its model card and license for the underlying model.
Training dataset: Aiwensile2/StorySeed. Check its current license and derivative-use restrictions before redistributing or commercializing adapted weights. In particular, a NoDerivatives dataset license can raise separate questions from the base model’s license; the base model’s license alone does not resolve those questions.

Author

Shashank Vats · Hugging Face · GitHub