Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,30 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
tags:
|
| 3 |
+
- about
|
| 4 |
+
---
|
| 5 |
+
|
| 6 |
+
# David Aylward
|
| 7 |
+
|
| 8 |
+
AI tinkerer in South Africa. I take language models apart on two 8GB gaming GPUs to
|
| 9 |
+
see what's inside, then put smaller ones back together. Everything I break and
|
| 10 |
+
everything I learn ships public: weights, measurements, mistakes and all.
|
| 11 |
+
|
| 12 |
+
## Current projects
|
| 13 |
+
|
| 14 |
+
| | |
|
| 15 |
+
|---|---|
|
| 16 |
+
| **[Qwen3.8-Whittle-16B](https://huggingface.co/logic65/Qwen3.8-Whittle-16B)** | A 27B whittled to 16.8B with a logit lens and a pricing table, healed with one A100 evening. 36/39 on the field battery at 20 tok/s on consumer GPUs. Research preview, v2. |
|
| 17 |
+
| **[The un-repaired cut](https://huggingface.co/logic65/Qwen3.8-p44w75-16.8B-unrepaired)** | The raw surgery artifact: full measurement history, every pricing run, every script. The damage profile is the science. |
|
| 18 |
+
|
| 19 |
+
## How this works
|
| 20 |
+
|
| 21 |
+
No lab, no cluster. Measurements run on a Ryzen 2700X with an RTX 4060 and an
|
| 22 |
+
RTX 3050. Training runs on rented A100 hours. Every experiment publishes its wins,
|
| 23 |
+
its bugs, and its dead ends on the model cards themselves.
|
| 24 |
+
|
| 25 |
+
## Support the tinkering
|
| 26 |
+
|
| 27 |
+
**[☕ ko-fi.com/davida81328](https://ko-fi.com/davida81328)**
|
| 28 |
+
|
| 29 |
+
Every donation becomes A100 hours, and every A100 hour ends up as a public model
|
| 30 |
+
or a public measurement.
|