Full weight fine tuned on two epochs of SlimOrca. Uses Mistral Instruct's prompt format.
The base model for this came from a variation on Undi's Mistral 11B recipe. The o_proj and down_proj tensors were set to zero in the added layers, making the output exactly identical to Mistral 7B before training.
Benchmarks look good locally but still evaluating actual usefulness.
Update: this turned out great! 10/10 would recommend as a training approach.
Reproducing
This mergekit config was used to produce the base model:
slices:
- sources:
- model: mistralai/Mistral-7B-v0.1
layer_range: [0, 24]
- sources: # add middle layers with residuals scaled to zero
- model: mistralai/Mistral-7B-v0.1
layer_range: [8, 24]
parameters:
scale:
- filter: o_proj
value: 0.0
- filter: down_proj
value: 0.0
- value: 1.0
- sources:
- model: mistralai/Mistral-7B-v0.1
layer_range: [24, 32]
merge_method: passthrough
dtype: bfloat16
The axolotl config for fine tuning is available here.
Open LLM Leaderboard Evaluation Results
Detailed results can be found here
| Metric | Value |
|---|---|
| Avg. | 66.12 |
| AI2 Reasoning Challenge (25-Shot) | 64.25 |
| HellaSwag (10-Shot) | 83.81 |
| MMLU (5-Shot) | 63.66 |
| TruthfulQA (0-shot) | 54.66 |
| Winogrande (5-shot) | 77.98 |
| GSM8k (5-shot) | 52.39 |