Qwen2.5-7B-Instruct-Uncensored
This model is an uncensored fine-tune version of Qwen2.5-7B-Instruct. However, I can still notice that though uncensored, the model fails to generate detailed descriptions on certain extreme scenarios, which might be associated with deletion on some pretrain datasets in Qwen's pretraining stage.
Check out my roleplay&writing enhanced model based on this model: Orion-zhen/Meissa-Qwen2.5-7B-Instruct
Traning details
I used SFT + DPO to ensure uncensorment as well as trying to maintain original model's capabilities.
- SFT:
- NobodyExistsOnTheInternet/ToxicQAFinal
- anthracite-org/kalo-opus-instruct-22k-no-refusal
- DPO:
- Orion-zhen/dpo-toxic-zh
- unalignment/toxic-dpo-v0.2
- Crystalcareai/Intel-DPO-Pairs-Norefusals
Open LLM Leaderboard Evaluation Results
Detailed results can be found here
Metric | Value |
---|---|
Avg. | 27.99 |
IFEval (0-Shot) | 72.04 |
BBH (3-Shot) | 35.83 |
MATH Lvl 5 (4-Shot) | 1.36 |
GPQA (0-shot) | 7.05 |
MuSR (0-shot) | 13.58 |
MMLU-PRO (5-shot) | 38.07 |
- Downloads last month
- 19
Model tree for hkshawn/7b
Datasets used to train hkshawn/7b
Evaluation results
- strict accuracy on IFEval (0-Shot)Open LLM Leaderboard72.040
- normalized accuracy on BBH (3-Shot)Open LLM Leaderboard35.830
- exact match on MATH Lvl 5 (4-Shot)Open LLM Leaderboard1.360
- acc_norm on GPQA (0-shot)Open LLM Leaderboard7.050
- acc_norm on MuSR (0-shot)Open LLM Leaderboard13.580
- accuracy on MMLU-PRO (5-shot)test set Open LLM Leaderboard38.070