You can deploy this model and start earning money today!

RichardErkhov/
oobabooga_-_CodeBooga-34B-v0.1-gguf

Name: oobabooga_-_CodeBooga-34B-v0.1-gguf
Availability: InStock

endpoints_compatible

gguf

template

Views

No views yet

Quantization made by Richard Erkhov.

Github

Discord

Request more models

CodeBooga-34B-v0.1 - GGUF

Model creator: https://huggingface.co/oobabooga/
Original model: https://huggingface.co/oobabooga/CodeBooga-34B-v0.1/

Name	Quant method	Size
CodeBooga-34B-v0.1.Q2_K.gguf	Q2_K	11.65GB
CodeBooga-34B-v0.1.IQ3_XS.gguf	IQ3_XS	12.93GB
CodeBooga-34B-v0.1.IQ3_S.gguf	IQ3_S	13.65GB
CodeBooga-34B-v0.1.Q3_K_S.gguf	Q3_K_S	13.6GB
CodeBooga-34B-v0.1.IQ3_M.gguf	IQ3_M	14.18GB
CodeBooga-34B-v0.1.Q3_K.gguf	Q3_K	15.19GB
CodeBooga-34B-v0.1.Q3_K_M.gguf	Q3_K_M	15.19GB
CodeBooga-34B-v0.1.Q3_K_L.gguf	Q3_K_L	16.55GB
CodeBooga-34B-v0.1.IQ4_XS.gguf	IQ4_XS	16.99GB
CodeBooga-34B-v0.1.Q4_0.gguf	Q4_0	17.74GB
CodeBooga-34B-v0.1.IQ4_NL.gguf	IQ4_NL	17.92GB
CodeBooga-34B-v0.1.Q4_K_S.gguf	Q4_K_S	17.87GB
CodeBooga-34B-v0.1.Q4_K.gguf	Q4_K	18.83GB
CodeBooga-34B-v0.1.Q4_K_M.gguf	Q4_K_M	18.83GB
CodeBooga-34B-v0.1.Q4_1.gguf	Q4_1	19.69GB
CodeBooga-34B-v0.1.Q5_0.gguf	Q5_0	21.64GB
CodeBooga-34B-v0.1.Q5_K_S.gguf	Q5_K_S	21.64GB
CodeBooga-34B-v0.1.Q5_K.gguf	Q5_K	22.2GB
CodeBooga-34B-v0.1.Q5_K_M.gguf	Q5_K_M	22.2GB
CodeBooga-34B-v0.1.Q5_1.gguf	Q5_1	23.59GB
CodeBooga-34B-v0.1.Q6_K.gguf	Q6_K	25.78GB
CodeBooga-34B-v0.1.Q8_0.gguf	Q8_0	33.39GB

Original model description:

license: llama2

CodeBooga-34B-v0.1

This is a merge between the following two models:

It was created with the BlockMerge Gradient script, the same one that was used to create MythoMax-L2-13b, and with the same settings. The following YAML was used:

yaml

model_path1: "Phind_Phind-CodeLlama-34B-v2_safetensors"
model_path2: "WizardLM_WizardCoder-Python-34B-V1.0_safetensors"
output_model_path: "CodeBooga-34B-v0.1"
operations:
  - operation: lm_head # Single tensor
    filter: "lm_head"
    gradient_values: [0.75]
  - operation: embed_tokens # Single tensor
    filter: "embed_tokens"
    gradient_values: [0.75]
  - operation: self_attn
    filter: "self_attn"
    gradient_values: [0.75, 0.25]
  - operation: mlp
    filter: "mlp"
    gradient_values: [0.25, 0.75]
  - operation: layernorm
    filter: "layernorm"
    gradient_values: [0.5, 0.5]
  - operation: modelnorm # Single tensor
    filter: "model.norm"
    gradient_values: [0.75]

Prompt format

Both base models use the Alpaca format, so it should be used for this one as well.

Below is an instruction that describes a task. Write a response that appropriately completes the request.

### Instruction:
Your instruction

### Response:
Bot reply

### Instruction:
Another instruction

### Response:
Bot reply

Evaluation

(This is not very scientific, so bear with me.)

I made a quick experiment where I asked a set of 3 Python and 3 Javascript questions (real world, difficult questions with nuance) to the following models:

This one
A second variant generated with model_path1 and model_path2 swapped in the YAML above, which I called CodeBooga-Reversed-34B-v0.1
WizardCoder-Python-34B-V1.0
Phind-CodeLlama-34B-v2

Specifically, I used 4.250b EXL2 quantizations of each. I then sorted the responses for each question by quality, and attributed the following scores:

4th place: 0
3rd place: 1
2nd place: 2
1st place: 4

The resulting cumulative scores were:

CodeBooga-34B-v0.1: 22
WizardCoder-Python-34B-V1.0: 12
Phind-CodeLlama-34B-v2: 7
CodeBooga-Reversed-34B-v0.1: 1

CodeBooga-34B-v0.1 performed very well, while its variant performed poorly, so I uploaded the former but not the latter.

Quantized versions

GGUF

TheBloke has kindly provided GGUF quantizations for llama.cpp:

https://huggingface.co/TheBloke/CodeBooga-34B-v0.1-GGUF

RichardErkhov/oobabooga_-_CodeBooga-34B-v0.1-gguf