Phi-3 Model lineage (mlx_lm.models.phi3)

[uogbuji]

Any notes, considerations, etc. around models derived from Microsoft Phi-3

[uogbuji]

Very interesting model: Triplex: a SOTA LLM for knowledge graph construction.

Knowledge graphs, like Microsoft's Graph RAG, enhance RAG methods but are expensive to build. Triplex offers a 98% cost reduction for knowledge graph creation, outperforming GPT-4 at 1/60th the cost and enabling local graph building with SciPhi's R2R.

Triplex is a finetuned version of Phi3-3.8B for creating knowledge graphs from unstructured data developed by SciPhi.AI. It works by extracting triplets - simple statements consisting of a subject, predicate, and object - from text or other data sources.

It really Wants to just generate triples, though! 😂

python -m mlx_lm.generate --model SciPhi/Triplex --prompt "Write a story about elephants" --temp 0.5 --max-tokens 512         
…
Special tokens have been added in the vocabulary, make sure the associated word embeddings are fine-tuned or trained.
==========
Prompt: <s><|user|>
Write a story about elephants<|end|>
<|assistant|>

```json
{
    "entities_and_triples": []
}
    ```

Worth more pondering at some point…

[uogbuji]

Phi3 with the quick_check. Weirdly in one of the square root ones did the correct tool-calling, and then decided to gnerate some unnecessary Python code.

❯ python test/quick_check.py mlx-community/Phi-3-mini-128k-instruct-4bit
Special tokens have been added in the vocabulary, make sure the associated word embeddings are fine-tuned or trained.
Model type: phi3
<|assistant|> Hello! I'm just a digital assistant, so I don't have feelings, but I'm functioning optimally and ready to help you. How can I assist you today?
 ======================================== Country extraction
<|assistant|> As per the given sentence, 'Adamma went home to Nigeria for the hols', the country mentioned is Nigeria. Here is the answer in the requested JSON format:

```json
[
    {
        "name": "Nigeria",
        "continent": "Africa"
    }
]

======================================== Square root of 256, pt 1
⚙️ Calling tool square_root with args {'square': 256}
⚙️ Tool call result: 16.0
<|assistant|> The square root of 256 is 16.
======================================== Square root of 256, pt 2
⚙️ Calling tool square_root with args {'square': 256}
⚙️ Tool call result: 16.0
<|assistant|> ```python

import math

def square_root(number):

return math.sqrt(number)

result = square_root(256)

print(result)

 ======================================== Usain bolt

 ======================================== END CHECK

[uogbuji]

Tried mlx-community/Phi-3.5-mini-instruct-bf16, newly released. It seems to wander into incoherence a lot, in addition to the insertion of instruct-speak which I'm seeing with all the Phi3 arch models.

[uogbuji]

Hey phi-4 is out. Seems to have weird try-hard tendencies, though.

❯ python test/quick_check.py mlx-community/phi-4-4bit
Model type: phi3
Hello! I'm here and ready to assist you. How can I help you today?<|im_end|>

 ======================================== Country extraction
[{"name": "Nigeria", "continent": "Africa"}]

 ======================================== Square root of 256, pt 1
The square root of 256 is 16. This is because 16 multiplied by itself (16 × 16) equals 256. Therefore, the positive square root of 256 is 16.<|im_end|>

 ======================================== Square root of 256, pt 2
The square root of 256 is 16.

 ======================================== Square root of 256, pt 3
The square root of 256 is 16. This is because 16 multiplied by itself (16 × 16) equals 256. Therefore, the positive square root of 256 is 16.<|im_end|>

 ======================================== Usain bolt
To find Usain Bolt's average velocity during his 100m race, we use the formula for average velocity:

\[
\text{Average velocity} = \frac{\text{Total distance}}{\text{Total time}}
\]

Given:
- Total distance = 100 meters
- Total time = 9.58 seconds

Substituting these values into the formula gives:

\[
\text{Average velocity} = \frac{100 \, \text{m}}{9.58 \, \text{s}} \approx 10.44 \, \text{m/s}
\]

Therefore, Usain

 ======================================== END CHECK