LongformerRM-Unison

LongformerRM-Unison

Absolute multi-head reward model for literary writing.

Heads

This model outputs 3 logits in this order:

1. style
2. faith
3. identifier

Apply sigmoid to each logit to obtain scores in [0, 1].

Intended use

This model can be used in two ways:

1. Rewrite scoring (primary use)

Score a (source passage, rewrite) pair for:

• stylistic quality
• semantic/content faithfulness
• identifier preservation (names, dates, numbers, protected spans)

2. Standalone chunk scoring (experimental workaround)

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 meaningful
• faith_score and identifier_score become proxy scores relative to the synthetic pseudo-source
• the resulting overall score is a proxy chunk-ranking score, not a true rewrite-faithfulness score

Important

This is not a comparative Bradley-Terry reward model.
It is an absolute scorer. Score each candidate independently, then sort externally.

Input format

Use exactly:

### Original Draft:
{prompt}

### Rewritten Version:
{response}

For Longformer inference, set global_attention_mask[:, 0] = 1.

Rewrite scoring

For normal rewrite evaluation, use the real source passage as prompt and the candidate rewrite as response.

Standalone chunk scoring

If you only have a passage chunk and no source passage, the recommended workaround is:

• generate a synthetically corrupted, flatter, still grammatical version of the chunk
• place that synthetic corruption in Original Draft
• place the real chunk in Rewritten Version

This better matches the model’s training format than an empty prompt.

Caveat

In standalone chunk mode, faith and identifier are not true faithfulness metrics. They only measure agreement with the synthetic corrupted prompt.

Recommended composite score for rewrite scoring

overall_score = style_score * (0.5 * faith_score + 0.5 * identifier_score) * (identifier_score ** 1.5)

Recommended score for standalone chunk scoring

You can rank by:

• proxy_overall_score if using a synthetic corrupted prompt
• or just style_score if you want the simplest signal

The synthetic-prompt method usually produces more separation between chunks than an empty-prompt style-only setup.

Output head order

[style_logit, faith_logit, identifier_logit]

Base model

allenai/longformer-base-4096