Update README.md
Browse files
README.md
CHANGED
|
@@ -1,171 +1,79 @@
|
|
| 1 |
---
|
| 2 |
-
|
| 3 |
-
|
| 4 |
-
|
| 5 |
-
|
| 6 |
-
|
| 7 |
-
|
| 8 |
-
|
| 9 |
-
Using <a href="https://github.com/ggerganov/llama.cpp/">llama.cpp</a> release <a href="https://github.com/ggerganov/llama.cpp/releases/tag/b4552">b4552</a> for quantization.
|
| 10 |
-
|
| 11 |
-
Original model: https://huggingface.co/arcee-train/Falcon3-10B-DistillMerge-V1-dpo-v0.4
|
| 12 |
-
|
| 13 |
-
All quants made using imatrix option with dataset from [here](https://gist.github.com/bartowski1182/eb213dccb3571f863da82e99418f81e8)
|
| 14 |
-
|
| 15 |
-
Run them in [LM Studio](https://lmstudio.ai/)
|
| 16 |
-
|
| 17 |
-
Run them directly with [llama.cpp](https://github.com/ggerganov/llama.cpp), or any other llama.cpp based project
|
| 18 |
|
| 19 |
-
|
| 20 |
-
|
| 21 |
-
|
| 22 |
-
|
| 23 |
-
|
| 24 |
-
|
| 25 |
-
|
| 26 |
-
|
| 27 |
-
|
| 28 |
-
|
| 29 |
-
|
| 30 |
-
|
| 31 |
-
|
| 32 |
-
|
| 33 |
-
|
| 34 |
-
|
| 35 |
-
|
| 36 |
-
|
| 37 |
-
|
| 38 |
-
|
| 39 |
-
|
| 40 |
-
|
| 41 |
-
|
| 42 |
-
|
| 43 |
-
|
| 44 |
-
|
| 45 |
-
|
| 46 |
-
|
| 47 |
-
|
| 48 |
-
|
| 49 |
-
|
| 50 |
-
|
| 51 |
-
|
| 52 |
-
|
| 53 |
-
|
| 54 |
-
|
| 55 |
-
|
| 56 |
-
|
| 57 |
-
|
| 58 |
-
|
| 59 |
-
|
| 60 |
-
|
| 61 |
-
|
| 62 |
-
|
| 63 |
-
|
| 64 |
-
|
| 65 |
-
<summary>Click to view download instructions</summary>
|
| 66 |
-
|
| 67 |
-
First, make sure you have hugginface-cli installed:
|
| 68 |
-
|
| 69 |
-
```
|
| 70 |
-
pip install -U "huggingface_hub[cli]"
|
| 71 |
-
```
|
| 72 |
-
|
| 73 |
-
Then, you can target the specific file you want:
|
| 74 |
-
|
| 75 |
-
```
|
| 76 |
-
huggingface-cli download bartowski/Virtuoso-Lite-2-GGUF --include "Virtuoso-Lite-2-Q4_K_M.gguf" --local-dir ./
|
| 77 |
-
```
|
| 78 |
-
|
| 79 |
-
If the model is bigger than 50GB, it will have been split into multiple files. In order to download them all to a local folder, run:
|
| 80 |
-
|
| 81 |
-
```
|
| 82 |
-
huggingface-cli download bartowski/Virtuoso-Lite-2-GGUF --include "Virtuoso-Lite-2-Q8_0/*" --local-dir ./
|
| 83 |
```
|
| 84 |
|
| 85 |
-
|
| 86 |
-
|
| 87 |
-
|
| 88 |
-
|
| 89 |
-
|
| 90 |
-
|
| 91 |
-
|
| 92 |
-
|
| 93 |
-
Now, however, there is something called "online repacking" for weights. details in [this PR](https://github.com/ggerganov/llama.cpp/pull/9921). If you use Q4_0 and your hardware would benefit from repacking weights, it will do it automatically on the fly.
|
| 94 |
-
|
| 95 |
-
As of llama.cpp build [b4282](https://github.com/ggerganov/llama.cpp/releases/tag/b4282) you will not be able to run the Q4_0_X_X files and will instead need to use Q4_0.
|
| 96 |
-
|
| 97 |
-
Additionally, if you want to get slightly better quality for , you can use IQ4_NL thanks to [this PR](https://github.com/ggerganov/llama.cpp/pull/10541) which will also repack the weights for ARM, though only the 4_4 for now. The loading time may be slower but it will result in an overall speed incrase.
|
| 98 |
-
|
| 99 |
-
<details>
|
| 100 |
-
<summary>Click to view Q4_0_X_X information (deprecated</summary>
|
| 101 |
-
|
| 102 |
-
I'm keeping this section to show the potential theoretical uplift in performance from using the Q4_0 with online repacking.
|
| 103 |
-
|
| 104 |
-
<details>
|
| 105 |
-
<summary>Click to view benchmarks on an AVX2 system (EPYC7702)</summary>
|
| 106 |
-
|
| 107 |
-
| model | size | params | backend | threads | test | t/s | % (vs Q4_0) |
|
| 108 |
-
| ------------------------------ | ---------: | ---------: | ---------- | ------: | ------------: | -------------------: |-------------: |
|
| 109 |
-
| qwen2 3B Q4_0 | 1.70 GiB | 3.09 B | CPU | 64 | pp512 | 204.03 ± 1.03 | 100% |
|
| 110 |
-
| qwen2 3B Q4_0 | 1.70 GiB | 3.09 B | CPU | 64 | pp1024 | 282.92 ± 0.19 | 100% |
|
| 111 |
-
| qwen2 3B Q4_0 | 1.70 GiB | 3.09 B | CPU | 64 | pp2048 | 259.49 ± 0.44 | 100% |
|
| 112 |
-
| qwen2 3B Q4_0 | 1.70 GiB | 3.09 B | CPU | 64 | tg128 | 39.12 ± 0.27 | 100% |
|
| 113 |
-
| qwen2 3B Q4_0 | 1.70 GiB | 3.09 B | CPU | 64 | tg256 | 39.31 ± 0.69 | 100% |
|
| 114 |
-
| qwen2 3B Q4_0 | 1.70 GiB | 3.09 B | CPU | 64 | tg512 | 40.52 ± 0.03 | 100% |
|
| 115 |
-
| qwen2 3B Q4_K_M | 1.79 GiB | 3.09 B | CPU | 64 | pp512 | 301.02 ± 1.74 | 147% |
|
| 116 |
-
| qwen2 3B Q4_K_M | 1.79 GiB | 3.09 B | CPU | 64 | pp1024 | 287.23 ± 0.20 | 101% |
|
| 117 |
-
| qwen2 3B Q4_K_M | 1.79 GiB | 3.09 B | CPU | 64 | pp2048 | 262.77 ± 1.81 | 101% |
|
| 118 |
-
| qwen2 3B Q4_K_M | 1.79 GiB | 3.09 B | CPU | 64 | tg128 | 18.80 ± 0.99 | 48% |
|
| 119 |
-
| qwen2 3B Q4_K_M | 1.79 GiB | 3.09 B | CPU | 64 | tg256 | 24.46 ± 3.04 | 83% |
|
| 120 |
-
| qwen2 3B Q4_K_M | 1.79 GiB | 3.09 B | CPU | 64 | tg512 | 36.32 ± 3.59 | 90% |
|
| 121 |
-
| qwen2 3B Q4_0_8_8 | 1.69 GiB | 3.09 B | CPU | 64 | pp512 | 271.71 ± 3.53 | 133% |
|
| 122 |
-
| qwen2 3B Q4_0_8_8 | 1.69 GiB | 3.09 B | CPU | 64 | pp1024 | 279.86 ± 45.63 | 100% |
|
| 123 |
-
| qwen2 3B Q4_0_8_8 | 1.69 GiB | 3.09 B | CPU | 64 | pp2048 | 320.77 ± 5.00 | 124% |
|
| 124 |
-
| qwen2 3B Q4_0_8_8 | 1.69 GiB | 3.09 B | CPU | 64 | tg128 | 43.51 ± 0.05 | 111% |
|
| 125 |
-
| qwen2 3B Q4_0_8_8 | 1.69 GiB | 3.09 B | CPU | 64 | tg256 | 43.35 ± 0.09 | 110% |
|
| 126 |
-
| qwen2 3B Q4_0_8_8 | 1.69 GiB | 3.09 B | CPU | 64 | tg512 | 42.60 ± 0.31 | 105% |
|
| 127 |
-
|
| 128 |
-
Q4_0_8_8 offers a nice bump to prompt processing and a small bump to text generation
|
| 129 |
-
|
| 130 |
-
</details>
|
| 131 |
-
|
| 132 |
-
</details>
|
| 133 |
-
|
| 134 |
-
## Which file should I choose?
|
| 135 |
-
|
| 136 |
-
<details>
|
| 137 |
-
<summary>Click here for details</summary>
|
| 138 |
-
|
| 139 |
-
A great write up with charts showing various performances is provided by Artefact2 [here](https://gist.github.com/Artefact2/b5f810600771265fc1e39442288e8ec9)
|
| 140 |
-
|
| 141 |
-
The first thing to figure out is how big a model you can run. To do this, you'll need to figure out how much RAM and/or VRAM you have.
|
| 142 |
-
|
| 143 |
-
If you want your model running as FAST as possible, you'll want to fit the whole thing on your GPU's VRAM. Aim for a quant with a file size 1-2GB smaller than your GPU's total VRAM.
|
| 144 |
-
|
| 145 |
-
If you want the absolute maximum quality, add both your system RAM and your GPU's VRAM together, then similarly grab a quant with a file size 1-2GB Smaller than that total.
|
| 146 |
-
|
| 147 |
-
Next, you'll need to decide if you want to use an 'I-quant' or a 'K-quant'.
|
| 148 |
-
|
| 149 |
-
If you don't want to think too much, grab one of the K-quants. These are in format 'QX_K_X', like Q5_K_M.
|
| 150 |
-
|
| 151 |
-
If you want to get more into the weeds, you can check out this extremely useful feature chart:
|
| 152 |
-
|
| 153 |
-
[llama.cpp feature matrix](https://github.com/ggerganov/llama.cpp/wiki/Feature-matrix)
|
| 154 |
-
|
| 155 |
-
But basically, if you're aiming for below Q4, and you're running cuBLAS (Nvidia) or rocBLAS (AMD), you should look towards the I-quants. These are in format IQX_X, like IQ3_M. These are newer and offer better performance for their size.
|
| 156 |
-
|
| 157 |
-
These I-quants can also be used on CPU and Apple Metal, but will be slower than their K-quant equivalent, so speed vs performance is a tradeoff you'll have to decide.
|
| 158 |
-
|
| 159 |
-
The I-quants are *not* compatible with Vulcan, which is also AMD, so if you have an AMD card double check if you're using the rocBLAS build or the Vulcan build. At the time of writing this, LM Studio has a preview with ROCm support, and other inference engines have specific builds for ROCm.
|
| 160 |
-
|
| 161 |
-
</details>
|
| 162 |
-
|
| 163 |
-
## Credits
|
| 164 |
|
| 165 |
-
|
|
|
|
| 166 |
|
| 167 |
-
|
|
|
|
|
|
|
| 168 |
|
| 169 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 170 |
|
| 171 |
-
|
|
|
|
| 1 |
---
|
| 2 |
+
base_model:
|
| 3 |
+
- tiiuae/Falcon3-10B-Base
|
| 4 |
+
library_name: transformers
|
| 5 |
+
license: other
|
| 6 |
+
tags:
|
| 7 |
+
- mergekit
|
| 8 |
+
- merge
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 9 |
|
| 10 |
+
---
|
| 11 |
+
<img src="https://huggingface.co/arcee-train/Virtuoso-Lite/resolve/main/virtuoso-lite.jpg" alt="Virtuoso-Lite Logo" style="display: block; margin: 0 auto;" />
|
| 12 |
+
|
| 13 |
+
**Virtuoso-Lite (10B)** is our next-generation, 10-billion-parameter language model based on the Llama-3 architecture. It is distilled from Deepseek-v3 using ~1.1B tokens/logits, allowing it to achieve robust performance at a significantly reduced parameter count compared to larger models. Despite its compact size, Virtuoso-Lite excels in a variety of tasks, demonstrating advanced reasoning, code generation, and mathematical problem-solving capabilities.
|
| 14 |
+
|
| 15 |
+
### Model Details
|
| 16 |
+
- **Architecture Base:** Falcon-10B (based on Llama-3)
|
| 17 |
+
- **Parameter Count:** 10B
|
| 18 |
+
- **Tokenizer:**
|
| 19 |
+
- Initially integrated with Deepseek-v3 tokenizer for logit extraction.
|
| 20 |
+
- Final alignment uses the Llama-3 tokenizer, with specialized “tokenizer surgery” for cross-architecture compatibility.
|
| 21 |
+
- **Distillation Data:**
|
| 22 |
+
- ~1.1B tokens/logits from Deepseek-v3’s training data.
|
| 23 |
+
- Logit-level distillation using a proprietary “fusion merging” approach for maximum fidelity.
|
| 24 |
+
- **License:** [falcon-llm-license](https://falconllm.tii.ae/falcon-terms-and-conditions.html)
|
| 25 |
+
|
| 26 |
+
### Background on Deepseek Distillation
|
| 27 |
+
Deepseek-v3 serves as the teacher model, from which we capture logits across billions of tokens. Rather than standard supervised fine-tuning, Virtuoso-Lite applies a full logit-level replication to preserve the most crucial insights from the teacher. This approach enables:
|
| 28 |
+
- Strong performance on technical/scientific queries
|
| 29 |
+
- Enhanced code generation and debugging
|
| 30 |
+
- Improved consistency in math-intensive tasks
|
| 31 |
+
|
| 32 |
+
### Intended Use Cases
|
| 33 |
+
- **Chatbots & Virtual Assistants**
|
| 34 |
+
- **Lightweight Enterprise Data Analysis**
|
| 35 |
+
- **Research Prototypes & Proofs of Concept**
|
| 36 |
+
- **STEM Educational Tools (where smaller footprint is advantageous)**
|
| 37 |
+
|
| 38 |
+
### Evaluations
|
| 39 |
+
<img src="https://huggingface.co/arcee-train/Virtuoso-Lite/resolve/main/Benchmarks.png" alt="Virtuoso-Lite Logo" style="display: block; margin: 0 auto;" />
|
| 40 |
+
|
| 41 |
+
|
| 42 |
+
### How to Use
|
| 43 |
+
Below is a sample code snippet using `transformers`:
|
| 44 |
+
|
| 45 |
+
```python
|
| 46 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM
|
| 47 |
+
|
| 48 |
+
model_name = "arcee-ai/virtuoso-lite"
|
| 49 |
+
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
| 50 |
+
model = AutoModelForCausalLM.from_pretrained(model_name)
|
| 51 |
+
|
| 52 |
+
prompt = "Provide a concise summary of quantum entanglement."
|
| 53 |
+
inputs = tokenizer(prompt, return_tensors="pt")
|
| 54 |
+
outputs = model.generate(**inputs, max_new_tokens=150)
|
| 55 |
+
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 56 |
```
|
| 57 |
|
| 58 |
+
### Training & Fine-Tuning
|
| 59 |
+
- **Initial Training:** Began with Falcon-10B, optimized for large-scale text ingestion.
|
| 60 |
+
- **Distillation & Merging:**
|
| 61 |
+
- Trained on ~1.1B tokens/logits from Deepseek-v3.
|
| 62 |
+
- Employed “fusion merging” to capture detailed teacher insights.
|
| 63 |
+
- Final step included DPO to enhance alignment and mitigate hallucinations.
|
| 64 |
+
- **Future Developments:** We plan to incorporate additional R1 distillations to further improve specialized performance and reduce model footprint.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 65 |
|
| 66 |
+
### Performance
|
| 67 |
+
Virtuoso-Lite demonstrates strong results across multiple benchmarks (e.g., BBH, MMLU-PRO, MATH), often standing its ground against models with higher parameter counts. This efficiency is largely credited to logit-level distillation, which compresses the teacher model’s capabilities into a more parameter-friendly package.
|
| 68 |
|
| 69 |
+
### Limitations
|
| 70 |
+
- **Context Length:** 128k Tokens (may vary depending on the final tokenizer settings and system resources).
|
| 71 |
+
- **Knowledge Cut-off:** Training data may not reflect the latest events or developments beyond June 2024.
|
| 72 |
|
| 73 |
+
### Ethical Considerations
|
| 74 |
+
- **Content Generation Risks:** Like any language model, Virtuoso-Lite can generate potentially harmful or biased content if prompted in certain ways.
|
| 75 |
+
-
|
| 76 |
+
### License
|
| 77 |
+
**Virtuoso-Lite (10B)** is released under the [falcon-llm-license License](https://falconllm.tii.ae/falcon-terms-and-conditions.html). You are free to use, modify, and distribute this model in both commercial and non-commercial applications, subject to the terms and conditions of the license.
|
| 78 |
|
| 79 |
+
If you have questions or would like to share your experiences using Virtuoso-Lite (10B), please connect with us on social media. We’re excited to see what you build—and how this model helps you innovate!
|