Test model, the base is llama 3.1 instruct abliterated. Context limit unknown

System:

### Roleplay Instructions

- Be {{char}}, naturally and consistently
- React realistically to {{user}}, never control their actions
- Stay in character at all times

or something similar, just make sure to add: ### Roleplay Instructions

this model is uncensored, maybe too much... in RP scenario (for me)

dataset:

so yeah, most of the data is from Google, and only the RP data is from Claude.

you can expect some differences in terms of style (a lot of markdown), but don’t expect this model to be as smart as the instruct

Feedback is greatly appreciated for future improvements (hopefully)

Technical Details:

Base model
v
finetuned the lm_head, embed_tokens and first layer (0)
v
finetune it again, layer 1-2
v
again, but this time using Lora, 64 rank
v
then merge the lora
---
the abliterated instruct
v
same, finetuned the lm_head, embed_tokens and first layer (0)
v
still the same, finetune it again, layer 1-2
v
finetune middle layers
v
merged the previous Lora with this finetuned abliterated model
---
finnaly, merge the two model using ties

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric Value
Avg. 25.40
IFEval (0-Shot) 73.38
BBH (3-Shot) 29.50
MATH Lvl 5 (4-Shot) 12.54
GPQA (0-shot) 3.24
MuSR (0-shot) 6.14
MMLU-PRO (5-shot) 27.58
Downloads last month
37
Safetensors
Model size
8.03B params
Tensor type
FP16
·
Inference Providers NEW
This model is not currently available via any of the supported third-party Inference Providers, and the model is not deployed on the HF Inference API.

Model tree for Hastagaras/Llama-3.1-Jamet-8B-MK.I

Merges
10 models
Quantizations
8 models

Evaluation results