The AskBeforeAnswer model is a clarification-seeking language model
based on Qwen2.5-7B-Instruct.
Instead of immediately answering an ambiguous question, the model is trained
to determine whether clarification is required and, when necessary, identify
the missing information and ask a targeted clarification question.
Production Model
This repository contains the exact W&B artifact approved by the project's
model promotion procedure.
The model was obtained directly from the immutable W&B artifact recorded in
the promotion record.
The DVC-managed training output is not modified during deployment.
Behavior
The model produces structured responses containing:
Action: whether to Clarify or Answer
Reasoning: why clarification is or is not required
Facets: missing information required to disambiguate the question
Response: either a clarification question or a direct answer
Expected format:
text
1Action: Clarify|Answer
2Reasoning: <reasoning>
3Facets: <list of missing facets>
4Response: <clarifying question or direct answer>
Training
The production winner was selected through the project's experiment
selection, verification, and promotion procedure.
The training pipeline supports:
Supervised Fine-Tuning (SFT)
Direct Preference Optimization (DPO)
Group Relative Policy Optimization (GRPO)
Odds Ratio Preference Optimization (ORPO)
This repository corresponds specifically to the model variant recorded in
the promotion record.
Evaluation
Evaluation results are maintained in the project evaluation artifacts.
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
base_model_name = "Qwen/Qwen2.5-7B-Instruct"
adapter_model_name = "chrisjcc/ask-before-answer"
tokenizer = AutoTokenizer.from_pretrained(
base_model_name,
)
model = AutoModelForCausalLM.from_pretrained(
base_model_name,
)
model = PeftModel.from_pretrained(
model,
adapter_model_name,
)
Dataset
The training datasets are available from:
chrisjcc/ask-before-answer-dataset
The repository contains separate sft and dpo configurations.
Intended Use
This model is intended for research into clarification-seeking behavior in
language models, particularly for systems that should distinguish between
answerable and underspecified user questions.
Limitations
The model may incorrectly classify questions as ambiguous or unambiguous.
Its generated reasoning and answers should not be treated as authoritative.
The model was trained on English-language data and may not generalize
reliably to other languages or domains.
Reproducibility
The model published here was selected through the project's experiment
selection, verification, and promotion procedure.
The promotion record identifies the exact W&B artifact version and digest
used for deployment.
The local DVC training artifact is treated as immutable during deployment.
