ablation
Collection
Eklav passage-reranking ablation & baseline checkpoints: answer-only SFT and the mask-only loss-masking ablation. • 3 items • Updated
Eklav-Reranker-AnswerOnly-GLM-Z1-9B
LoRA adapter for passage reranking, trained with the answer-only SFT
method on base model THUDM/GLM-Z1-9B-0414, from the Eklav project.
Method
Answer-only SFT is a lean baseline: the prompt carries no reasoning hint, and
the training target is just the bare verdict (true/false) with no
<think>...</think> chain-of-thought at all. It is trained on the full
(short) response with no loss masking, since there is nothing to mask.
This is not the same as the eklav-mask-only ablation, which keeps the
full reasoning trace in the response and masks it from the loss instead of
omitting it from the data. See
AdarshSingh7647/Eklav-Reranker-MaskOnly-Qwen3-8B
for that ablation.
Training data
AdarshSingh7647/Eklav-Reranker-AnswerOnly-Data
Checkpoint
Final checkpoint, step 5968 (training complete).
Usage
LoRA adapter (PEFT format, r=32, alpha=64) using the glmz1 chat template.
Load with the base model:
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
base = AutoModelForCausalLM.from_pretrained("THUDM/GLM-Z1-9B-0414")
model = PeftModel.from_pretrained(base, "AdarshSingh7647/Eklav-Reranker-AnswerOnly-GLM-Z1-9B")
tokenizer = AutoTokenizer.from_pretrained("AdarshSingh7647/Eklav-Reranker-AnswerOnly-GLM-Z1-9B")
Part of the Eklav project.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
Model tree for AdarshSingh7647/Eklav-Reranker-AnswerOnly-GLM-Z1-9B
Base model
zai-org/GLM-Z1-9B-0414