Language: English | 中文

InnoSpark3.1-27B-1008

InnoSpark3.1-27B-1008 is a 27B education-focused model built on Qwen3.8-27B through reinforcement learning (RL) post-training. It targets educational question answering, learning support, and educational assistance, including concept explanation, guided problem solving, personalized tutoring, and teaching material generation.

Model Details

Item Description
Model name InnoSpark3.1-27B-1008
Base model Qwen3.8-27B
Parameters 27B
Post-training method Reinforcement learning (RL)
Post-training focus Educational QA, learning support, educational assistance
Primary use cases Educational QA, learning support, educational assistance
Evaluated benchmarks EduBench, LongTutor

RL Post-Training

InnoSpark3.1-27B-1008 undergoes RL post-training on Qwen3.8-27B for educational question answering, learning support, and educational assistance.

Evaluation

EduBench: Dimension Scores

Dimension Qwen3.8-27B InnoSpark3.1-27B-1008 Change
Instruction Following & Task Completion 9.608 9.842 +0.233
Content Relevance & Scope Control 9.715 9.907 +0.192
Scenario Element Integration 8.434 9.230 +0.795
Basic Factual Accuracy 9.522 9.833 +0.312
Domain Knowledge Accuracy 9.049 9.567 +0.518
Reasoning Process Rigor 9.434 9.808 +0.373
Clarity, Simplicity & Inspiration 8.788 9.053 +0.265
Higher-Order Thinking & Skill Development 8.332 9.385 +1.053
Personalization, Adaptation & Learning Support 9.375 9.957 +0.582
Role & Tone Consistency 9.008 9.622 +0.613

* EduBench primarily evaluates three task types: IP, PLS, and TMG.

LongTutor

Metric Qwen3.8-27B InnoSpark3.1-27B-1008 Change
Overall memory accuracy 87.20% 88.93% +1.73 pp
Information extraction accuracy 97.80% 98.80% +1.00 pp
Multi-session reasoning accuracy 72.90% 74.90% +2.00 pp
Hallucination check accuracy 90.90% 93.10% +2.20 pp
Diagnosis accuracy 42.20% 45.50% +3.30 pp
Diagnosis Macro-F1 0.3738 0.4262 +0.0524
History utilization 3.828 4.588 +0.760
Strategy alignment 4.352 4.908 +0.556
Coherence 4.311 4.776 +0.465
Comprehension 4.058 4.450 +0.392
Overall teaching quality 4.038 4.688 +0.650

* LongTutor is evaluated on the human-annotated XES3G5M Gold set.

Usage

Transformers

import torch
from transformers import AutoProcessor, Qwen3_5ForConditionalGeneration

model_id = "sii-research/InnoSpark3.1-27B-1008"

processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
model = Qwen3_5ForConditionalGeneration.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
    trust_remote_code=True,
)

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "text",
                "text": "请用适合初中生的方式讲解一元二次方程的求根公式,并给出一个例题。",
            }
        ],
    }
]

inputs = processor.apply_chat_template(
    messages,
    add_generation_prompt=True,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=1024)
response = processor.decode(
    outputs[0][inputs["input_ids"].shape[-1]:],
    skip_special_tokens=True,
)
print(response)

vLLM

vllm serve sii-research/InnoSpark3.1-27B-1008 --trust-remote-code

Intended Use

  • Educational QA: explain concepts, answer subject-specific questions, and discuss solutions.
  • Learning support: guide problem solving through questions, use learning history, and adapt explanations to the learner.
  • Educational assistance: support lesson planning, teaching material generation, and instructional strategy suggestions.

Limitations

Like other large language models, InnoSpark3.1-27B-1008 may generate inaccurate, incomplete, or biased content. Outputs in educational settings should be reviewed by qualified educators when used for high-stakes learning, assessment, or student guidance. The model should not be used as the sole source for factual verification, grading decisions, psychological counseling, medical advice, legal advice, or other safety-critical decisions.

Evaluation results may vary with prompt format, decoding parameters, evaluation implementation, and data version. Users should conduct additional evaluations for their specific use cases before deploying the model in production or classroom environments.

Contributors

Contributor Link
songsiyu Google Scholar
luye GitHub, Google Scholar

License

See LICENSE for the Apache-2.0 license included with this checkpoint.

Citation

@misc{innospark3_1_27b_1008,
  title = {InnoSpark3.1-27B-1008},
  author = {songsiyu and luye},
  year = {2026}
}
Downloads last month
36
Safetensors
Model size
28B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for sii-research/InnoSpark3.1-27B-1008

Quantizations
2 models

Collection including sii-research/InnoSpark3.1-27B-1008