KO | EN

GitLite

Minimalist Open Source Repository Search Engine

gitlite — search

#dsh #ai-agents #dsh-plugin #codex #developer-tools #claude-code #windows #desktop-app #typescript #rust #ai-agent #cli

TinyEngram

★ 1,043 Open GitHub ↗

Research of DeepSeek Engram Architecture based on Qwen-3 and Stable Diffusion series.

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# TinyEngram: Exploring New Axis of Scaling and Memory Injection

<p align="center">
  <strong>Open research on DeepSeek-AI's Engram and memory injection in Qwen, StableDiffusion and more.</strong>
</p>

[**Report Issues**](https://github.com/AutoArk/TinyEngram/issues)

</div>

> [!NOTE]
> **TL;DR:** TinyEngram demonstrates that Engram-based memory injection outperforms LoRA in both **parameter efficiency** and **catastrophic forgetting resistance**—and extends seamlessly to **vision** (e.g., Stable Diffusion) for lightweight, composable concept injection. All code, logs, and experiments are open!
>
> *If you find TinyEngram useful, a ⭐ helps support the project.*

<br>

<div align="center">
<details open>
<summary style="
      cursor: pointer; 
      display: inline-flex; 
      align-items: center; 
      justify-content: center;
      font-weight: 700; 
      font-size: 1.15em; 
      color: #24292e; 
      margin-bottom: 20px;">
      <span style="margin-right: 8px;">📢</span> Latest Announcements
    </summary>

- <span style="color: #6a737d; font-family: monospace;">2026.05.20</span> — 📝 **TinyEngram-Vision technical report is ready.** We organized the vision findings into a complete technical report to invite discussion and further exploration. [Read the report](./doc/paper/tinyengram_vision_paper.pdf), or check out the [arXiv version](https://arxiv.org/abs/2605.20309) if you prefer.
- <span style="color: #6a737d; font-family: monospace;">2026.02.12</span> — 🖼️ **TinyEngram meets Vision!** We injected visual concepts into Stable Diffusion through Engram, check our new cross-modal experiment!
- <span style="color: #6a737d; font-family: monospace;">2026.02.02</span> — 📌 Released reproduction scripts for **Engram vs LoRA** experiment.
- <span style="color: #6a737d; font-family: monospace;">2026.01.30</span> — 📌 Added comparison of **catastrophic forgetting** between **TinyEngram and LoRA**.
- <span style="color: #6a737d; font-family: monospace;">2026.01.30</span> — 📌 Added **parameter ablation studies** of TinyEngram with convergence observations.
- <span style="color: #6a737d; font-family: monospace;">2026.01.23</span> — 🎉 Initial TinyEngram commit.

</div>
</details>
</div>

<br>

<div align="center">
  <p style="font-size: 0.95em; color: #586069; margin-bottom: 8px;">
    🔍 Quick Navigation
  </p>
  <p>
    <a href="#-key-finding-1-engram-as-parameter-efficient-fine-tuning-method">Key Finding 1: Engram as Parameter Efficient Fine-Tuning Method</a><br/>
    <a href="#-key-finding-2-engram-outperforms-lora-in-catastrophic-forgetting">Key Finding 2: Engram Outperforms LoRA in Catastrophic Forgetting</a><br/>
    <a href="#vision-engram">TinyEngram-Vision: Engram Goes Multimodal</a>
  </p>
</div>

## ⚙️ Quick Environment Setup

TinyEngram provides a clean, pinned direct-dependency file for training and
vision reproduction:

```bash
conda create -n tinyengram python=3.10 -y
conda activate tinyengram
pip install --upgrade pip
pip install -r requirements.txt
```

For CUDA notes and optional evaluation dependencies, see
[doc/reproduction/environment.md](./doc/reproduction/environment.md).

## 📖 Introduction

**TinyEngram** is an open research project exploring the [Engram](https://github.com/deepseek-ai/Engram) architecture—an LLM enhancement that boosts phrase-level understanding by integrating a compact N-gram memory module and a gated retrieval mechanism into key transformer layers.

Built on [Qwen](https://github.com/QwenLM/Qwen), **TinyEngram** provides a lightweight, ready-to-train codebase for anyone to reproduce, experiment with, or extend Engram-style models. We actively share new experiments, training logs, and findings right here—making this repo both a toolkit and a living research notebook.

Beyond LLMs, we propose a modality-agnostic memory architecture. Engram in Stable Diffusion serves as one instantiation, proving that non-textual concepts can be retrieved and integrated as efficiently as language.

> [!TIP]
> **Join the Research**
> You are welcome to propose any questions in the [Issues](https://github.com/AutoArk/TinyEngram/issues). We will burn our own GPUs to research on any interesting questions. Join us in evolving how LLMs remember what matters! 🧠✨

<span id="vision-engram"></span>
## 🖼️ TinyEngram-Vision: Engram Goes Multimodal

> **Technical Report:** We have organized the TinyEngram-Vision findings into a complete technical report. [Read the report](./doc/paper/tinyengram_vision_paper.pdf) for the full methodology, experiments, and conclusions, or check out the [arXiv version](https://arxiv.org/abs/2605.20309) if you prefer.

**Can Engram's memory mechanism work beyond text?**

We extended TinyEngram to **Stable Diffusion**, treating visual concepts as "memories" capable of being injected into the Text Encoder. By simply recognizing specific N-grams in the prompt, we inject learned embeddings that guide the generation—**all without fine-tuning the massive U-Net or DiT backbone.**

It's a lightweight, composable way to "teach" the model new subjects (like specific characters) while keeping the original weights frozen.

**Why is this cool?**
*   **Minimal & Surgical**: We construct a minimal Engram vocabulary specifically for your target phrase.
*   **Infinite Composability**: Since Engram relies on exact N-gram matching (hard hash collisions), **memories strictly do not interfere with each other.** You can stack thousands of different character/style engrams together, and they will only trigger when their exact name is called—**zero degradation** to the base model's general capabilities.

<div align="center">
  <img src="doc/experiments/figures/vision_tinyengram_intro.png" width="80%" alt="vision_engram_intro"/>
  <p><i>Information Injection in Text-to-Image via Engram (Death Stranding is the best game.)</i></p>
</div>

🔗 [**Click here to view the full report (SD1.5 & SD3.5 experiments)**](./doc/experiments/experimental_report_tinyengram_vision.md)

### Reproduce our experiments

Interested in injecting your own concepts (or even your cat 🐱) into Stable Diffusion? 
👉 **[Check out the reproduction guide here](./doc/reproduction/reproduce_tinyengram_vision.md)**

We provide everything needed to get started:
- Training scripts for **SD1.5** & **SD3.5**
- Pre-processed datasets
- Inference demos

> **Have Fun!**

## 🧪 Key Finding 1. Engram as Parameter Efficient Fine-Tuning Method

### 1. Engram works as Parameter Efficient Fine-Tuning Method

<table align="center">
  <tr>
    <td align="center">
      <img src="doc/train.png" height="260" style="object-fit: contain;" />
      <br/>
      <sub>Training Setup</sub>
    </td>
  </tr>
</table>

We insert several Engram modules into decoder layers of Qwen. We fine-tune the Engram module on a subset of the [Biomed-Enriched](https://huggingface.co/datasets/almanach/Biomed-Enriched) dataset. Only added parameters are trainable during the fine-tuning.

The train and eval loss  demonstrate robust convergence. This confirms that the Engram module effectively learns specialized biomedical knowledge while preserving the stability of the underlying pre-trained knowledge base.

<table>
  <tr>
    <td align="center">
      <img src="doc/experiments/figures/sft_train.png"
           height="260"
           style="object-fit: contain;" />
      <br/>
      <sub>Training Loss</sub>
    </td>
    <td align="center">
      <img src="doc/experiments/figures/sft_eval.png"
           height="260"
           style="object-fit: contain;" />
      <br/>
      <sub>Validation Performance</sub>
    </td>
  </tr>
</table>

| Biomedical Task                  | Qwen3-0.6B | Engram SFT |
|----------------------|------------------------|----------------|
| MMLU_Clinical Knowledge   | 0.3358                 | 0.4415         |
| MMLU_Medical Genetics     | 0.3700                 | 0.4400         |
| MMLU_Prof. Medicine       | 0.3199                 | 0.4559         |
| PubMedQA             | 0.5700                 | 0.6250         |

### 2. Catastrophic Forgetting

*   **Objective**: Verify if integrating Engram memory harms the model's pre-trained general capabilities while adapting to new domains.
*   **Methodology**: We fine-tune the [Biomed-Enriched](https://huggingface.co/datasets/almanach/Biomed-Enriched) on Qwen, and evaluate the checkpoint on general benchmarks (We evaluate on MMLU, excluding all biomedical-related subtasks).
*   **Full Results**: [Click here to view detailed results](./doc/experiments/catastrophic_forgetting.md)

| Task Group        | Qwen 3-0.6B | Engram SFT   |
|-------------------|----------------|-------------------------------|
| mmlu (overall)| 0.4034           | 0.4500 (⬆️ +0.0466)           |
| humanities    | 0.4433             | 0.4691 (⬆️ +0.0258)           |
| other         | 0.4271                    | 0.4696 (⬆️ +0.0425)           |
| social sciences| 0.4826                   | 0.5389 (⬆️ +0.0563)           |
| stem          | 0.3508                    | 0.4088 (⬆️ +0.0580)           |

> **📌 Update (2026.01.30):** We have added a new set of experiments comparing Engram and LoRA on catastrophic forgetting. Please refer to [**Engram vs LoRA Catastrophic Forgetting Experiment**](#1-engram-vs-lora-catastrophic-forgetting-experiment) for details.

### 3. Vocabulary Scalability Analysis

*   **Objective**: Investigate the relationship between Engram memory size (vocabulary size) and performance gains.
*   **Methodology**: Train multiple models with varying `engram_vocab_size` (e.g., 2k vs 10k vs 20k vs 100k) and observe the impact on biomedical validation loss.
*   **Full Results**: Larger representation capacities do not necessarily translate into better performance.
In our experiments, we observe an apparent trade-off: smaller capacities may suffer from semantic collisions, while larger ones can become difficult to fully utilize given limited data. [Click here to view detailed results](./doc/experiments/engram_scaling_on_small_dataset.md)

| Task                  | Nano (2k/0.2k) | Small (10k/1k) | Medium (20k/2k) | Large (100k/10k) | Qwen3-0.6B (Baseline) | Winner             |
|----------------------|----------------|----------------|------------------|-------------------|------------------------|------------------|
| MMLU_Clinical Knowledge   | 0.3736         | 0.4415         | 0.4302           | 0.4226            | 0.3358                 | Small 🏆         |
| MMLU_Medical Genetics     | 0.3900         | 0.4400         | 0.4400           | 0.4100            | 0.3700                 | Small/Med 🤝     |
| MMLU_Prof. Medicine       | 0.4081         | 0.4559         | 0.4228           | 0.4412            | 0.3199                 | Small 🏆         |
| PubMedQA             | 0.6240         | 0.6250         | 0.6170           | 0.6150            | 0.5700                 | Small 🏆         |

> **📌 Update (2026.01.30):**  
  We have expanded our study with a comprehensive ablation of TinyEngram’s configurable hyperparameters. Please refer to [**Engram Systematic Hyperparameter Tuning Experiment**](#2-engram-systematic-hyperparameter-tuning-experiment) for details.

### Reproduce our experiments

To reproduce the experiments conducted in **Key Finging 1**, please refer to [this guide.](./doc/reproduction/reproduce_exp.md)

## 🧪 Key Finding 2. Engram Outperforms LoRA in Catastrophic Forgetting

LoRA is the de-facto PEFT method, So how does Engram compare? We also conduct systematic hyperparameter tuning to understand Engram better.

### 1. Engram vs LoRA Catastrophic Forgetting Experiment

**Preliminary observation:** In our experiments, Engram shows noticeably better resistance to catastrophic forgetting than LoRA.

| Model Architecture | Adaptation Metric (Eval Loss) $\downarrow$ | General Capability (TruthfulQA MC1) $\uparrow$ | General Capability (TruthfulQA MC2) $\uparrow$ | $\Delta$ (MC2 vs Base) |
| :--- | :--- | :--- | :--- | :--- |
| **Qwen-0.6B (Base)** | N/A | 0.2583 | 0.4269 | - |
| **LoRA (Rank 16)** | 0.1862 | 0.2485 | 0.4078 | <span style="color:red">-1.91%</span> |
| **TinyEngram** | **0.1850** | **0.2644** | **0.4340** | <span style="color:green">+0.71%</span> |

> It is worth noting that LoRA generally converges faster. In our experiments, LoRA could reach an even lower loss (0.1458) quickly, but the trade-off was severe: catastrophic forgetting worsened significantly ($MC1: 0.2472$, $MC2: 0.3993$). Engram provides a safer learning path.

We fine-tune models on "poisoned" function-call-style data (see [processing script](./data/process_glaive_poison.py)) based on the [glaive-function-calling-v2](https://huggingface.co/datasets/glaiveai/glaive-function-calling-v2) dataset, which encourages a strong bias toward structured function-call outputs. We then evaluate both LoRA and Engram on TruthfulQA, a natural language QA benchmark, to examine how well they retain general-language capabilities under this distribution shift. [Click here to view detailed results.](./doc/experiments/catastrophic_forgetting_vs_lora.md)

### 2. Engram Systematic Hyperparameter Tuning Experiment

During initial trials, we observed that **LoRA converges faster** than the default Engram configuration. To enable a scientifically sound comparison, we conducted a systematic hyperparameter study to calibrate Engram such that it reaches **evaluation loss levels comparable to LoRA** on the same training data.

Using the small-scale, filtered [glaive-function-calling-v2](https://huggingface.co/datasets/glaiveai/glaive-function-calling-v2) dataset, we ablated key Engram parameters beyond vocabulary size, including:  
- **N-gram order**
- **Vocabulary size** 
- **Embedding dimension per n-gram**
- **Number of hash heads per n-gram**
- **Target layer(s) for Engram injection**

<div align="center">
  
  <img src="doc/experiments/figures/parameters_tuning/overview.png" width="80%" alt="overview_of_parameter_exp" />
  <p><i>Detailed analysis is available via the link below.</i></p>
</div>

We hope this experiment can serve as a solid starting point for parameter selection in similar small-scale supervised fine-tuning (SFT) scenarios.
🔗 [Click here to view detailed results.](./doc/experiments/engram_parameters_tuning.md)

### Reproduce our experiments

Reproduction details of experiments conducted in **Key Finging 2**: please refer to [this guide](./doc/reproduction/reproduce_lora_exp.md).

## 🗺️ More Research is on the way!
| Category | Item | Status |
| :--- | :--- | :---: |
| **Multimodal** | Stable Diffusion Injection | ✅ |
| **Engram as PEFT** | Engram works | ✅ |
| | Catastrophic Forgetting | ✅ |
| | Vocabulary Scalability | ✅ |
| | vs LoRA | ✅ |
| | Hyperparameter Tuning | ✅ |
| More | More | ⬜ |

## 🙏 Acknowledgements

We borrowed a lot of code from the following excellent projects:

- [Engram](https://github.com/deepseek-ai/Engram)
- [Qwen](https://github.com/QwenLM/Qwen)
- [lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness)
- [Diffusers](https://huggingface.co/docs/diffusers/en/index)

We thank the authors of training datasets that help our research:

- [Biomed-Enriched](https://huggingface.co/datasets/almanach/Biomed-Enriched)
- [glaive-function-calling-v2](https://huggingface.co/datasets/glaiveai/glaive-function-calling-v2)

## 🔗 Citation

If you find TinyEngram useful for your research or projects, please cite us:

```bibtex
@misc{cai2026tinyengramtriggerindexedconcepttables,
  title         = {Tiny-Engram: Trigger-Indexed Concept Tables for Generative Vision},
  author        = {Runyuan Cai and Yiming Wang and Yu Lin and Xiaodong Zeng},
  year          = {2026},
  eprint        = {2605.20309},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2605.20309}
}
```

TinyEngram

Explore Similar Repositories

Was this content helpful?

Recent Feedback

Download README