|
|
--- |
|
|
license: apache-2.0 |
|
|
tags: |
|
|
- unsloth |
|
|
- Uncensored |
|
|
- text-generation-inference |
|
|
- transformers |
|
|
- unsloth |
|
|
- llama |
|
|
- trl |
|
|
- roleplay |
|
|
- conversational |
|
|
datasets: |
|
|
- iamketan25/roleplay-instructions-dataset |
|
|
- N-Bot-Int/Iris-Uncensored-R1 |
|
|
- N-Bot-Int/Moshpit-Combined-R2-Uncensored |
|
|
- N-Bot-Int/Mushed-Dataset-Uncensored |
|
|
- N-Bot-Int/Muncher-R1-Uncensored |
|
|
- N-Bot-Int/Millia-R1_DPO |
|
|
language: |
|
|
- en |
|
|
base_model: |
|
|
- N-Bot-Int/MiniMaid-L1 |
|
|
pipeline_tag: text-generation |
|
|
library_name: peft |
|
|
metrics: |
|
|
- character |
|
|
--- |
|
|
 |
|
|
# MiniMaid-L2 |
|
|
- MiniMaid-L2 is a Finetuned Model of MiniMaid-L1 model, with even big and higher quality dataset used to generated roleplaying |
|
|
Capabilities, MiniMaid-L2 also were extracted from Knowledge Distilling A Popular Roleplaying Model named NoroMaid-7B-DPO, |
|
|
Which we've used to enchanced its lacking Ends for coherent And Good Roleplaying Capabilities. |
|
|
|
|
|
- MiniMaid-L2 Outcompete its predecessor as it uses a Clever Knowledge distilling to transfer Knowledge from NoroMaid, |
|
|
And Finetuned it, building on top of MiniMaid-L1 to Produce a better AI model. Sacrificing Some Non-noticable |
|
|
Token-Generation speed, with a near perfect and Competitive Model against **3b Alternatives**! |
|
|
|
|
|
# MiniMaid-L1 Base-Model Card Procedure: |
|
|
- **MiniMaid-L1** achieve a good Performance through process of DPO and Combined Heavy Finetuning, To Prevent Overfitting, |
|
|
We used high LR decays, And Introduced Randomization techniques to prevent the AI from learning and memorizing, |
|
|
However since training this on Google Colab is difficult, the Model might underperform or underfit on specific tasks |
|
|
Or overfit on knowledge it manage to latched on! However please be guided that we did our best, and it will improve as we move onwards! |
|
|
|
|
|
- MiniMaid-L2 is Another Instance of Our Smallest Model Yet! if you find any issue, then please don't hesitate to email us at: |
|
|
[nexus.networkinteractives@gmail.com](mailto:nexus.networkinteractives@gmail.com) |
|
|
about any overfitting, or improvements for the future Model **V3**, |
|
|
Once again feel free to Modify the LORA to your likings, However please consider Adding this Page |
|
|
for credits and if you'll increase its **Dataset**, then please handle it with care and ethical considerations |
|
|
|
|
|
- MiniMaid-L2 is |
|
|
- **Developed by:** N-Bot-Int |
|
|
- **License:** apache-2.0 |
|
|
- **Parent Model from model:** unsloth/llama-3.2-3b-instruct-unsloth-bnb-1bit |
|
|
- **Dataset Combined Using:** Mosher-R1(Propietary Software) |
|
|
|
|
|
- MiniMaid-L1 Official Metric Score |
|
|
 |
|
|
- Metrics Made By **ItsMeDevRoland** |
|
|
Which compares: |
|
|
- **MiniMaid-L1 GGUFF** |
|
|
- **MiniMaid-L2 GGUFF** |
|
|
Which are All Ranked with the Same Prompt, Same Temperature, Same Hardware(Google Colab), |
|
|
To Properly Showcase the differences and strength of the Models |
|
|
|
|
|
- **Visit Below to See details!** |
|
|
|
|
|
--- |
|
|
# 🧵 MiniMaid-L2: Small Size, Big Bite — The Next-Gen Roleplay Assistant |
|
|
> She’s sharper, deeper, and more immersive. And this time? She doesn’t just hold her own — she wins. |
|
|
|
|
|
 |
|
|
|
|
|
# MiniMaid-L2 builds on the scrappy L1 foundation and takes the lead over 3B giants like Hermes, Dolphin, and DeepSeek, with better consistency, longer outputs, and a massive boost to immersion. |
|
|
- 💬 Roleplay Evaluation (v1) |
|
|
- 🧠 Character Consistency: 0.84 |
|
|
- 🌊 Immersion: 0.47 |
|
|
-🧮 Overall RP Score: 0.76 |
|
|
- ✏️ Length Score: 1.00 |
|
|
- L2 scored +0.25 higher overall than L1, while beating top-tier 3B models in every major RP metric. |
|
|
|
|
|
|
|
|
# 📊 Efficient AND Smart |
|
|
- Inference Time: 54.2s — still 3x faster than Hermes |
|
|
- Tokens/sec: 6.88 — near-instant on consumer GPUs |
|
|
- BLEU/ROUGE-L: Stronger n-gram overlap than any 3B rival |
|
|
|
|
|
# MiniMaid-L2 shows that distilled models can outperform much larger ones — when trained right, even 1B can be the boss. |
|
|
- 🛠️ MiniMaid is Built For |
|
|
- High-fidelity RP generation |
|
|
- Lower-latency systems |
|
|
- Custom, character-driven storytelling |
|
|
|
|
|
> 🌱 L2 is the turning point — with upgraded conditioning, tighter personality anchoring, and narrative-aware outputs, she's evolving fast. |
|
|
|
|
|
“MiniMaid-L2 doesn’t just punch above her weight — she’s taking belts. A tighter model, a stronger performer, and still tiny enough to run on a toaster. RP just got smarter.” |
|
|
|
|
|
--- |
|
|
|
|
|
- # Notice |
|
|
- **For a Good Experience, Please use** |
|
|
- Low temperature 1.5, min_p = 0.1 and max_new_tokens = 128 |
|
|
|
|
|
|
|
|
- # Detail card: |
|
|
- Parameter |
|
|
- 1 Billion Parameters |
|
|
- (Please visit your GPU Vendor if you can Run 1B models) |
|
|
|
|
|
- Finetuning tool: |
|
|
- Unsloth AI |
|
|
- This llama model was trained 2x faster with [Unsloth](https://github.com/unslothai/unsloth) and Huggingface's TRL library. |
|
|
[<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>](https://github.com/unslothai/unsloth) |
|
|
- Fine-tuned Using: |
|
|
- Google Colab |
