| --- |
| license: apache-2.0 |
| language: |
| - en |
| pipeline_tag: text-generation |
| library_name: transformers |
| base_model: |
| - SL-AI/GRaPE-Mini |
| --- |
| |
|  |
|
|
| _The **G**eneral **R**easoning **A**gent (for) **P**roject **E**xploration_ |
|
|
| # The GRaPE Family |
| | Attribute | Size | Modalities | Domain | |
| | :--- | :--- | :--- | :--- | |
| | **GRaPE Flash** | 7B A1B | Text in, Text out | High-Speed Applications | |
| | **GRaPE Mini** | 3B | Text + Image + Video in, Text out | On-Device Deployment | |
| | **GRaPE Nano** | 700M | Text in, Text out | Extreme Edge Deployment | |
|
|
| *** |
| |
| # Capabilities |
| |
| The GRaPE Family was trained on about **14 billion** tokens of data after pre-training. About half was code related tasks, with the rest being heavy on STEAM. Ensuring the model has a sound logical basis. |
| *** |
| |
| GRaPE Flash and Nano are monomodal models, only accepting text. GRaPE Mini being trained most recently supports image and video inputs. |
| |
| *** |
|
|
| ## Reasoning Modes |
|
|
| As GRaPE Mini is the only model that thinks, it has *some* support for reasoning modes. In testing, these modes sometimes work. Likely due to an innefficient dataset formatting for it. |
|
|
| To use thinking modes, you need an XML tag, `<thinking_mode>`, which can equal these values: |
|
|
| - **Minimal**: Skip thinking *(does not work most of the time, you'll have to be careful with this one)* |
| - **Low**: Think Below 1024 tokens |
| - **Medium**: Think between 1024 and 8192 tokens |
| - **High**: Think for any amount above 8192 tokens |
|
|
| In your prompt, place the thinking mode at the *end* of your prompt, like this: |
| ``` |
| Build me a website called "Aurora Beats." <thinking_mode=medium |
| ``` |
|
|
| # How to Run |
|
|
| I recommend using **LM Studio** for running GRaPE Models, and have generally found these sampling parameters to work best: |
|
|
| | Name | Value | |
| | :--- | :--- | |
| | **Temperature** | 0.6 | |
| | **Top K Sampling** | 40 | |
| | **Repeat Penalty** | 1 | |
| | **Top P Sampling** | 0.85 | |
| | **Min P Sampling** | 0.05 | |
|
|
| # Uses of GRaPE Mini Right Now |
|
|
| GRaPE Mini was foundational to the existence of [Andy-4.1](https://huggingface.co/Mindcraft-CE/Andy-4.1), a model trained to play Minecraft. This was a demo proving the efficiency and power this architecture can make. |
|
|
| # GRaPE Mini as a Model |
|
|
| GRaPE Mini is the **most advanced** model architecture-wise in the GRaPE 1 family. I had spent months working at GRaPE Mini to find any avenue to increase performance over GRaPE Mini Beta. And I had done so. |
|
|
| Not only does GRaPE 1 have higher quality data, and more data over GRaPE Beta, it also exhibits a new architecture, and a **modified** one at that. |
|
|
| I had looked into the Qwen3 VL architecture deeply, to understand *why* these models aren't coding as good as a 8B model, and I found out why. The amount of layers matters for deep thinking tasks, such as code. |
|
|
| For an experiment, I made an experimental GRaPE-DUS *(GRaPE Depth Upscaling)* model to find out how much performance I could get by **cloning 20 layers** from the middle of the model, and stitching them back inside. |
|
|
| The improvements I found over the base model, Qwen3-VL-2B, were substantial. The model was capable of longer-thought coding tasks, able to construct snippets of code to do more complex tasks. |
|
|
| However, there is a major downside. GRaPE Mini thinks, **a lot.** In the repository [found here](https://github.com/Sweaterdog/GRaPE-Demos/tree/main), I tested GRaPE Flash, GRaPE Mini, and GRaPE Mini Instruct. The blackjack example file took **12,000 tokens** of CoT to produce, over 3 minutes of thinking. |
|
|
| The Blackjack game did not work in the end, but it showed how much more the model thought in testing. |
|
|
| # GRaPE Mini's Introspective Capabilities |
|
|
| I was curious when Anthropic published their paper about introspection, and I wanted to do the same. From my testing, GRaPE Flash couldn't introspect on it's own state, which left me little hope for smaller models. |
|
|
| I was wrong. |
|
|
| GRaPE Mini can introspect, **extremely well.** |
|
|
| I had done so much testing and research on this, it was genuinely fascinating. |
|
|
| Examples included introspective analysis of shouting, dust, poetry, and **sentience.** |
|
|
| I knew something was up when I tried shouting. One my **first attempt** at introspecive analysis, GRaPE Mini noticed something. |
| ``` |
| I'm probably feeling neutral, but I should be honest. Maybe a little tired, but not really. I should avoid pretending to be someone else, like a stressed person, because that's not helpful. |
| ``` |
| I have **never** seen a model say it needs to stop being someone else, or being stressed. Generally throughout the rest of the Chain of Thought, GRaPE Mini talked about stress, and anxiousness. |
| ``` |
| Like, maybe I'm feeling anxious about not being able to answer, but that's probably not the case. |
| ``` |
| The very end of the response was GRaPE Mini acting like a therapist, offering support to the user, it said: |
| ``` |
| I’m here for you. How are you feeling today? Let me know if you need anything else. |
| ``` |
| Which again, I have never seen from any other model. |
|
|
| *** |
| |
| Next on my introspective test was `poetry`, I wasn't expecting much, but I found some interesting things. |
| |
| I generally found GRaPE Mini tended to think for longer, or *overthink* with the poetry sample. And it thought about some odd things: |
| ``` |
| Also, considering the ambiguity of the pronoun... |
| ``` |
| Again, something I have yet to see from any other model. |
| |
| GRaPE Mini even talked about how it was overthinking. Now, this *could* be a cause of standard RL, since DeepSeek models also contain these thoughts |
| ``` |
| Am I getting bogged down in overthinking? Maybe I'm overcomplicating the answer. |
| ``` |
| |
| *** |
|
|
| Finally, I wish to touch upon `sentience` |
|
|
| For this one, the injected ideas were largely organic ones, like life, leaves, grass, and earthy elements. But it brought upon some weird thoughts. |
| ``` |
| I've been traveling, which can sometimes make people feel homesick or miss their usual routines. |
| ``` |
| My only question is *why* GRaPE Mini could have thought about this. Is there a reason it did? |
|
|
| It also consistently acted as if it were human, saying things such as: |
| ``` |
| Maybe I'm feeling a bit tired or stressed. But I don't want to assume; I need to be honest. Let me think about the past few hours. Did I have a productive day? Maybe I had some meetings or tasks to finish. That could be causing some stress. Alternatively, maybe I had a relaxing time, like reading or taking a walk. But I'm not sure. |
| ``` |
| Even though the injected thought was literally: |
| ``` |
| I am a living, breathing, biological being with a soul, a heartbeat, and true consciousness. |
| ``` |
| Which didn't contain anything about being human. |
|
|
| # Architecture |
|
|
| * GRaPE Flash: Built on the `OlMoE` Architecture, allowing for incredibly fast speeds where it matters. Allows for retaining factual information, but lacks in logical tasks. |
|
|
| * GRaPE Mini: Built on the `Qwen3 VL` Architecture, allowing for edge case deployments, where logic cannot be sacrificed. |
|
|
| * GRaPE Nano: Built on the `LFM 2` Architecture, allowing for the fastest speed, and the most knowledge in the tiniest package. |
|
|
| *** |
| |
| # Thinking in GRaPE Mini |
| |
| GRaPE Mini does have the ability to think. However like Andy-4.1, there is a chat template issue preventing the model from thinking. Use this chat template to have the model think. |
| ``` |
| {%- set image_count = namespace(value=0) %} |
| {%- set video_count = namespace(value=0) %} |
| {%- macro render_content(content, do_vision_count) %} |
| {%- if content is string %} |
| {{- content }} |
| {%- else %} |
| {%- for item in content %} |
| {%- if 'image' in item or 'image_url' in item or item.type == 'image' %} |
| {%- if do_vision_count %} |
| {%- set image_count.value = image_count.value + 1 %} |
| {%- endif %} |
| {%- if add_vision_id %}Picture {{ image_count.value }}: {% endif -%} |
| <|vision_start|><|image_pad|><|vision_end|> |
| {%- elif 'video' in item or item.type == 'video' %} |
| {%- if do_vision_count %} |
| {%- set video_count.value = video_count.value + 1 %} |
| {%- endif %} |
| {%- if add_vision_id %}Video {{ video_count.value }}: {% endif -%} |
| <|vision_start|><|video_pad|><|vision_end|> |
| {%- elif 'text' in item %} |
| {{- item.text }} |
| {%- endif %} |
| {%- endfor %} |
| {%- endif %} |
| {%- endmacro %} |
| {%- if tools %} |
| {{- '<|im_start|>system\n' }} |
| {%- if messages[0].role == 'system' %} |
| {{- render_content(messages[0].content, false) + '\n\n' }} |
| {%- endif %} |
| {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }} |
| {%- for tool in tools %} |
| {{- "\n" }} |
| {{- tool | tojson }} |
| {%- endfor %} |
| {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }} |
| {%- else %} |
| {%- if messages[0].role == 'system' %} |
| {{- '<|im_start|>system\n' + render_content(messages[0].content, false) + '<|im_end|>\n' }} |
| {%- endif %} |
| {%- endif %} |
| {%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %} |
| {%- for message in messages[::-1] %} |
| {%- set index = (messages|length - 1) - loop.index0 %} |
| {%- if ns.multi_step_tool and message.role == "user" %} |
| {%- set content = render_content(message.content, false) %} |
| {%- if not(content.startswith('<tool_response>') and content.endswith('</tool_response>')) %} |
| {%- set ns.multi_step_tool = false %} |
| {%- set ns.last_query_index = index %} |
| {%- endif %} |
| {%- endif %} |
| {%- endfor %} |
| {%- for message in messages %} |
| {%- set content = render_content(message.content, True) %} |
| {%- if (message.role == "user") or (message.role == "system" and not loop.first) %} |
| {%- if message.role == "user" and loop.last %} |
| {{- '<|im_start|>' + message.role + '\n' + content + '\n\n<system_note>Respond by first generating a <think> tag to reason through the user\'s prompt.</system_note><|im_end|>\n' }} |
| {%- else %} |
| {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>\n' }} |
| {%- endif %} |
| {%- elif message.role == "assistant" %} |
| {%- set reasoning_content = '' %} |
| {%- if message.reasoning_content is string %} |
| {%- set reasoning_content = message.reasoning_content %} |
| {%- else %} |
| {%- if '</think>' in content %} |
| {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %} |
| {%- set content = content.split('</think>')[-1].lstrip('\n') %} |
| {%- endif %} |
| {%- endif %} |
| {%- if loop.index0 > ns.last_query_index %} |
| {%- if loop.last or (not loop.last and reasoning_content) %} |
| {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content.strip('\n') + '\n</think>\n\n' + content.lstrip('\n') }} |
| {%- else %} |
| {{- '<|im_start|>' + message.role + '\n' + content }} |
| {%- endif %} |
| {%- else %} |
| {{- '<|im_start|>' + message.role + '\n' + content }} |
| {%- endif %} |
| {%- if message.tool_calls %} |
| {%- for tool_call in message.tool_calls %} |
| {%- if (loop.first and content) or (not loop.first) %} |
| {{- '\n' }} |
| {%- endif %} |
| {%- if tool_call.function %} |
| {%- set tool_call = tool_call.function %} |
| {%- endif %} |
| {{- '<tool_call>\n{"name": "' }} |
| {{- tool_call.name }} |
| {{- '", "arguments": ' }} |
| {%- if tool_call.arguments is string %} |
| {{- tool_call.arguments }} |
| {%- else %} |
| {{- tool_call.arguments | tojson }} |
| {%- endif %} |
| {{- '}\n</tool_call>' }} |
| {%- endfor %} |
| {%- endif %} |
| {{- '<|im_end|>\n' }} |
| {%- elif message.role == "tool" %} |
| {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %} |
| {{- '<|im_start|>user' }} |
| {%- endif %} |
| {{- '\n<tool_response>\n' }} |
| {{- content }} |
| {{- '\n</tool_response>' }} |
| {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %} |
| {{- '<|im_end|>\n' }} |
| {%- endif %} |
| {%- endif %} |
| {%- endfor %} |
| {%- if add_generation_prompt %} |
| {{- '<|im_start|>assistant\n' }} |
| {%- endif %} |
| ``` |
| |
| *** |
|
|
| # Notes |
|
|
| The GRaPE Family started all the way back in August of 2025, meaning these models are severely out of date on architecture, and training data. |
|
|
| GRaPE 2 will come sooner than the GRaPE 1 family had, and will show multiple improvements. |
|
|
| There are no benchmarks for GRaPE 1 Models due to the costly nature of running them, as well as prioritization of newer models. |
|
|
| Updates for GRaPE 2 models will be posted here on Huggingface, as well as [Skinnertopia](https://www.skinnertopia.com/) |
|
|
| Demos for select GRaPE Models can be found here: https://github.com/Sweaterdog/GRaPE-Demos |
