WIP, still updating. Lots of babbling and ramblings. Look at the pretty graphics instead.

Conclusions (WIP)

Upscaling can 'provide room' for further training.
Training upscaled models will result in retaining more of the original model's performance & behavior.
A 600MB dataset was nowhere near in stabilizing the empty/duplicated layers. (Pertubed rate of change remained the same for epoch 1 & 2)
(Not related to upscaling) The first two layers are sus - their weights are wildly different from the original. I wonder if we could recover smarts by merging that back in with base, or if those layers contain the most influence and must be preserved.

What is the 39B Upscale?

https://huggingface.co/TheSkullery/BA-Zephyria-39b

merge_method: passthrough
slices:
- sources:
  - layer_range: [0, 41]
    model: unsloth/Mistral-Small-Instruct-2409
- sources:
  - layer_range: [19, 41]
    model: unsloth/Mistral-Small-Instruct-2409
    parameters:
      scale:
      - filter: o_proj
        value: 0.0
      - filter: down_proj
        value: 0.0
      - value: 1.0
- sources:
  - layer_range: [19, 41]
    model: unsloth/Mistral-Small-Instruct-2409
    parameters:
      scale:
      - filter: o_proj
        value: 0.0
      - filter: down_proj
        value: 0.0
      - value: 1.0
- sources:
  - layer_range: [41, 55]
    model: unsloth/Mistral-Small-Instruct-2409

Layers 0 to 18 are original
Layers 19 to 41 are duplicated, zero'd out, and put in the middle twice
Layers 42 to 54 are original
down_proj and o_proj layers for the duplicated part have been nulled and will require healing to 'unignore' the added layers

[    Unique    ][    Duplicated    ][    Unique    ]
0 ----------- 18 19 ------------ 41 42 ---------- 54
     34.5%           41.8%            23.7%

How did the finetune go?

Weight Difference Visualization

Contorl A: Nemo x Rocinante
Control B: Small x Cydonia
Control C: Upscaled Nemo x Theia
Sample A: 39B Upscale x Tunguska 1 Epoch
Sample B: 39B Upscale x Tunguska 2 Epochs
Sample C: Tunguska 1 Epoch x Tunguska 2 Epochs

*Note the layer sequence and other labels since it will be unreadable for the 39B

*Visualization Issue: Some layer types like input_layernorm look unchanged because the highest value (usually layer 0) dilluted the entire heatmap

Control A (Nemo 12B & Rocinante 12B, similar training)

Control B (Small 22B & Cydonia 22B, similar training)

Control C (Upscaled Nemo 21B & Theia 21B, similar training)

Sample A: Tunguska 39B 1st Epoch vs. its 39B base

Sample B: Tunguska 39B 2nd Epoch vs. its 39B base

Sample C: Tunguska 39B 2nd Epoch vs. 1st Epoch

Glossary

WIP

Impressions

Using the same LR on a larger model should have been more destructive. In practice, the duplicated layers must have saved the upscale from becoming dumber than Cydonia. The larger model seems to have preserved more of its smarts thanks to the 'empty, extra' duplicate layers.
The upscale clearly has an effect on training. Control A & B show a smooth gradient unlike Control C & Sample A, B, C where the duplicated layers perturb the heatmap.
The duplicated layers seem to be playing catch up with the original layers. The change difference on the duplicate layers in Sample B show that they are more responsive to training than the original layers. There are no indications that the duplicate layers are 'slowing down', given that the gradient pattern remains the same in Sample A, B, and especially C (where C can be considered the rate of change).
In some layer types, the change difference for the first two layers are overwhelmingly huge compared to the other layers seeing as how they diluted the entire heatmap!
The duplicated layers on all layer types (except one) are extra sensitive. post_attention_layernorm interestingly had some changes in the upscale's duplicated layers, unlike Cydonia where latter layers were completely unchanged.
The duplicated layers in o_proj are less sensitive for some reason.

Eureka?

Top-to-Bottom Linear Gradient Observations

Take note of a few things
- Top layers = Ending layers (nearer to output)
- Bottom layers = Starting layers (nearer to input)
- Training a normal, non-upscaled model affects the top layers first and slowly descends to the bottom layers over time.
- Training an upscaled model with two slices of duplicate layers does two things:
  - Each slice of duplicated layers has its own gradient.
  - There's a 'ceiling value' for the duplicated layers in these slices.
  - Even when Tunguska's slices of duplicated layers are nearly saturated, the resulting model remains coherent and even performant.
Takeaways
- These slice of layers are more connected to each other than to the model's entirety.
  - [Question] Does this mean that the original layer before the slice is the one holding that whole duplicated slice together?
  - [Question] What if we interleave original and duplicate layers? Will that result in a more balanced, responsive upscale? (See Proposed Upscale Technique at the bottom)
- Saturating these duplicated layers MIGHT be a good goal to pursue.

Further Experimentation

Given how the duplicated layers seem to have a stabilizing effect, it begs the question:

What if we duplicate only ONE layer? What about five layers?

Will fewer empty layers dampen the stabilizing effect?
Will the few empty layers get 'filled' quickly? Will the 600MB dataset be enough?
Will there be a greater concetration of weight change in the duplicate layers? Or will it spill out to the original layers?

We've so far hypothesized that training 'slowly fills' the duplicated layers. If we intentionally undercook, will the duplicated layers look underfilled or can you fill it up with a few steps? In other words, can a single/few updates to the model reconnect the duplicated layers?

Are we really repairing the 'neurons' step-by-step, or have they been significantly rearranged by the first (few?) steps?
Or maybe this is false given the top-bottom gradient observation.

Proposed Upscale Technique

merge_method: passthrough
slices:
- sources:
  - layer_range: [0, 18]
    model: unsloth/Mistral-Small-Instruct-2409
# Original L19
- sources:
  - layer_range: [19, 19]
    model: unsloth/Mistral-Small-Instruct-2409
# Dupe A of L19
- sources:
  - layer_range: [19, 19]
    model: unsloth/Mistral-Small-Instruct-2409
    parameters:
      scale:
      - filter: o_proj
        value: 0.0
      - filter: down_proj
        value: 0.0
      - value: 1.0
# Dupe B of L19
- sources:
  - layer_range: [19, 19]
    model: unsloth/Mistral-Small-Instruct-2409
    parameters:
      scale:
      - filter: o_proj
        value: 0.0
      - filter: down_proj
        value: 0.0
      - value: 1.0
# Original L20
- sources:
  - layer_range: [20, 20]
    model: unsloth/Mistral-Small-Instruct-2409
# Dupe A of L20
- sources:
  - layer_range: [20, 20]
    model: unsloth/Mistral-Small-Instruct-2409
    parameters:
      scale:
      - filter: o_proj
        value: 0.0
      - filter: down_proj
        value: 0.0
      - value: 1.0
# Dupe B of L20
- sources:
  - layer_range: [20, 20]
    model: unsloth/Mistral-Small-Instruct-2409
    parameters:
      scale:
      - filter: o_proj
        value: 0.0
      - filter: down_proj
        value: 0.0
      - value: 1.0
# ... REPEAT UNTIL 41
- sources:
  - layer_range: [41, 55]
    model: unsloth/Mistral-Small-Instruct-2409

0 = original
X = duplicate

Previous Technique = 000000000000000000XXXXXXXXXXXXXXXXXXX0000000000
Proposed Technique = 00000000000XX0XX0XX0XX0XX0XX0XX0XX0XX0000000000