|
--- |
|
base_model: |
|
- TheDrummer/Anubis-70B-v1 |
|
- EVA-UNIT-01/EVA-LLaMA-3.33-70B-v0.1 |
|
- meta-llama/Llama-3.3-70B-Instruct |
|
- Sao10K/70B-L3.3-Cirrus-x1 |
|
- SicariusSicariiStuff/Negative_LLAMA_70B |
|
- Sao10K/L3.1-70B-Hanami-x1 |
|
library_name: transformers |
|
tags: |
|
- mergekit |
|
- merge |
|
license: llama3.3 |
|
--- |
|
|
|
[original model]: https://huggingface.co/Tarek07/Progenitor-V5-Final-LLaMa-70B |
|
|
|
## Update Feb 16, 2025 noon PST: |
|
|
|
Upon debugging with the author of the [original model], the author decided to redo the model as something is off with the weights. See discussion [here](https://huggingface.co/Tarek07/Progenitor-V5-Final-LLaMa-70B/discussions/1#67b24c8eb02f929c82a02a73). |
|
|
|
**NOTE: This means this repo's weights are also broken!** |
|
|
|
## Update Feb 16, 2025 morning PST: |
|
|
|
The author of the [original model] mentioned that this model gave very different outputs. See ongoing discussion [here](https://huggingface.co/Tarek07/Progenitor-V5-Final-LLaMa-70B/discussions/1#67b21fd3ba726eda5c98e812). |
|
|
|
# Overview |
|
|
|
The [original model] had invalid `tensor.Shape` for weights (`[1, 8192]`), raising following errors when loading with `transformers`: |
|
``` |
|
ValueError: Trying to set a tensor of shape torch.Size([1, 8192]) in "weight" (which has shape torch. Size ( [8192])), this looks incorrect. |
|
``` |
|
|
|
So I resized them into `[8192]` with following script: |
|
```python |
|
import os |
|
from safetensors.torch import load_file, save_file |
|
|
|
# Update this to point to your safetensors directory |
|
MODEL_DIR = "/root/.cache/huggingface/hub/models--Tarek07--Progenitor-V5-Final-LLaMa-70B/snapshots/8ca900fd3a65a725902d525e518be1bf374c0247" |
|
DEST_DIR = "/output/Progenitor-V5-Final-LLaMa-70B" |
|
|
|
def fix_shard(shard_path, output_path): |
|
# Load the shard |
|
data = load_file(shard_path) |
|
# data is a dict: key -> torch.Tensor |
|
|
|
# Go through every tensor and fix the shape if necessary |
|
for key, tensor in data.items(): |
|
# Check if the shape is (1, 8192) instead of (8192) |
|
if list(tensor.shape) == [1, 8192]: |
|
print(f" Fixing {key} in {os.path.basename(shard_path)} from {tensor.shape} to (8192,)") |
|
# Either squeeze(0) or view(-1) or view(8192): |
|
# data[key] = tensor.squeeze(0) |
|
# or |
|
data[key] = tensor.view(8192) |
|
|
|
# Save the fixed shard to output_path |
|
save_file(data, output_path, metadata={"format": "pt"}) |
|
print(f" -> Saved fixed shard to: {output_path}") |
|
|
|
def main(): |
|
# Look for .safetensors files in MODEL_DIR |
|
for filename in sorted(os.listdir(MODEL_DIR)): |
|
if filename.endswith(".safetensors"): |
|
shard_path = os.path.join(MODEL_DIR, filename) |
|
output_path = os.path.join(DEST_DIR, f"{filename}") |
|
|
|
print(f"Processing: {shard_path}") |
|
fix_shard(shard_path, output_path) |
|
|
|
if __name__ == "__main__": |
|
main() |
|
``` |
|
|
|
# Original README.md from here: |
|
|
|
This marks the culmination of my experiments with the Progenitor series. I fixed the typo I had earlier where it wasn't computing in float32, but 6 models in computed in float32 is a bit taxing on resources and time and so I left it for the configuration I thought was the best (it's not something I can afford to do with every model I make, just the worthwhile ones). This one also uses the Sicari's tokenizer which I find the best. |
|
# merge |
|
|
|
This is a merge of pre-trained language models created using [mergekit](https://github.com/cg123/mergekit). |
|
|
|
## Merge Details |
|
### Merge Method |
|
|
|
This model was merged using the [Linear DELLA](https://arxiv.org/abs/2406.11617) merge method using [meta-llama/Llama-3.3-70B-Instruct](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct) as a base. |
|
|
|
### Models Merged |
|
|
|
The following models were included in the merge: |
|
* [TheDrummer/Anubis-70B-v1](https://huggingface.co/TheDrummer/Anubis-70B-v1) |
|
* [EVA-UNIT-01/EVA-LLaMA-3.33-70B-v0.1](https://huggingface.co/EVA-UNIT-01/EVA-LLaMA-3.33-70B-v0.1) |
|
* [Sao10K/70B-L3.3-Cirrus-x1](https://huggingface.co/Sao10K/70B-L3.3-Cirrus-x1) |
|
* [SicariusSicariiStuff/Negative_LLAMA_70B](https://huggingface.co/SicariusSicariiStuff/Negative_LLAMA_70B) |
|
* [Sao10K/L3.1-70B-Hanami-x1](https://huggingface.co/Sao10K/L3.1-70B-Hanami-x1) |
|
|
|
### Configuration |
|
|
|
The following YAML configuration was used to produce this model: |
|
|
|
```yaml |
|
models: |
|
- model: Sao10K/L3.1-70B-Hanami-x1 |
|
parameters: |
|
weight: 0.20 |
|
density: 0.7 |
|
- model: Sao10K/70B-L3.3-Cirrus-x1 |
|
parameters: |
|
weight: 0.20 |
|
density: 0.7 |
|
- model: SicariusSicariiStuff/Negative_LLAMA_70B |
|
parameters: |
|
weight: 0.20 |
|
density: 0.7 |
|
- model: TheDrummer/Anubis-70B-v1 |
|
parameters: |
|
weight: 0.20 |
|
density: 0.7 |
|
- model: EVA-UNIT-01/EVA-LLaMA-3.33-70B-v0.1 |
|
parameters: |
|
weight: 0.20 |
|
density: 0.7 |
|
merge_method: della_linear |
|
base_model: meta-llama/Llama-3.3-70B-Instruct |
|
parameters: |
|
epsilon: 0.2 |
|
lambda: 1.1 |
|
int8_mask: true |
|
dtype: float32 |
|
out_dtype: bfloat16 |
|
tokenizer: |
|
source: SicariusSicariiStuff/Negative_LLAMA_70B |
|
``` |
|
|
