Stable LM 2 BRIEF 1.6B
Fine-tuned chat model based on stabilityai/stablelm-2-1_6b. It was trained for 21 epochs using 1024 context windows and a mixture of small subsets of UltraChat and OASST2. Since the model saw shorter dialogs it tends to be less verbose than StabilityAI's 1.6B chat model stabilityai/stablelm-2-zephyr-1_6b.
Instruction format:
<|user|>
Hi!<|endoftext|>
<|assistant|>
Hello there!<|endoftext|>
maxim-saplin/stablelm-2-brief-1_6b:
stabilityai/stablelm-2-zephyr-1_6b:
Available in 16-bit HF safetensors and 16 and 8-bit GGUF.
Training code is here
MT-Bench
########## First turn ##########
score
model turn
stablelm-2-brief-1_6b_2 1 3.240506
stablelm-2-brief-1_6b_3 1 3.202532
stablelm-2-brief-1_6b 1 2.850000*
########## Second turn ##########
score
model turn
stablelm-2-brief-1_6b_3 2 2.443038
stablelm-2-brief-1_6b_2 2 2.350000
stablelm-2-brief-1_6b 2 2.175000*
########## Average ##########
score
model
stablelm-2-brief-1_6b_3 2.822785
stablelm-2-brief-1_6b_2 2.792453
stablelm-2-brief-1_6b 2.512500*
stablelm-2-zephyr-1_6b scored 5.42