Absolute multi-head reward model for literary writing.
This model outputs 3 logits in this order:
stylefaithidentifierApply sigmoid to each logit to obtain scores in [0, 1].
This model can be used in two ways:
Score a (source passage, rewrite) pair for:
Score a single passage chunk by pairing it with a synthetically corrupted but grammatical pseudo-source derived from that same chunk.
In this mode:
style_score is still meaningfulfaith_score and identifier_score become proxy scores relative to the synthetic pseudo-sourceThis is not a comparative Bradley-Terry reward model.
It is an absolute scorer. Score each candidate independently, then sort externally.
Use exactly:
### Original Draft:
{prompt}
### Rewritten Version:
{response}
For Longformer inference, set global_attention_mask[:, 0] = 1.
For normal rewrite evaluation, use the real source passage as prompt and the candidate rewrite as response.
If you only have a passage chunk and no source passage, the recommended workaround is:
Original DraftRewritten VersionThis better matches the model’s training format than an empty prompt.
In standalone chunk mode, faith and identifier are not true faithfulness metrics. They only measure agreement with the synthetic corrupted prompt.
overall_score = style_score * (0.5 * faith_score + 0.5 * identifier_score) * (identifier_score ** 1.5)
You can rank by:
proxy_overall_score if using a synthetic corrupted promptstyle_score if you want the simplest signalThe synthetic-prompt method usually produces more separation between chunks than an empty-prompt style-only setup.
[style_logit, faith_logit, identifier_logit]
allenai/longformer-base-4096
model.safetensors 595 MBtraining_args.bin 5.8 kB