Compare the deal term to the draft: LiTiL Financing Review 9B
A focused financing-review model that compares an agreed term with the corresponding draft provision.
Give the model one agreed business term and the corresponding draft provision. Its comparison can then say what changed and cite both sections.
LiTiL Financing Review 9B handles that focused comparison. It takes a term-sheet excerpt and a draft-document excerpt with section labels, then returns a short Markdown issues memo with a proposed correction.
The model can compare liquidation preference, notice periods, option-pool size, board composition, and similar deal terms. It is most useful after another part of the system has already found and paired the relevant language.
Where it fits
The workflow starts with documents, but this model should not receive the whole closing set at once.
A practical stack does the following:
- Parse the term sheet and transaction documents.
- Extract and normalize the agreed deal terms.
- Find the draft provision that corresponds to each term.
- Send one labeled comparison to LiTiL Financing Review 9B.
- Display the memo beside the source language for review.
The model's job is step four. Retrieval and section pairing stay upstream. Citation checking and the decision to revise stay downstream.
That separation makes errors easier to diagnose. If the wrong provision was paired, fix retrieval. If the right provisions were supplied but the comparison was wrong, fix or replace the comparison model. If the memo was right but the proposed language did not match current drafting policy, fix the playbook.
What the output looks like
The published example compares a 1x non-participating liquidation preference in a term sheet with a 2x participating preference in a draft charter. The adapter identifies the mismatch, cites Term sheet §1 and draft charter §2, and proposes revising the draft to match the agreed 1x non-participating term.
The prompt asks for a memo of at most 100 words. That constraint keeps the result close to an issues list that a deal team can scan beside the source text.
For new comparisons, preserve the section labels and keep each prompt to one deal term. The runner also preserves the literal role strings used during training and stops at the next role boundary so the application receives one clean memo.
What was trained
The release is a PEFT LoRA adapter for Qwen/Qwen3.5-9B. Supervised tuning used 88 synthetic financing examples and 25 optimizer steps on Apple Silicon.
The training messages were reproduced from the authored synthetic generator, which does not read private source input. No private client or user data appears in the reviewed post-training set.
The release is about 116 MB and requires the separately available Qwen3.5-9B base model. The packaged examples let a team inspect recorded outputs before loading either set of weights.
What the measurements say
A fresh four-case diagnostic checked liquidation preference, notice periods, option-pool size, and a matching board provision. The adapter returned the correct core comparison in all four first responses.
A focused liquidation-preference follow-up produced the complete correction in 57 generated tokens. Both arms used the same excerpts and role-boundary stops; the matched-base final-answer comparison remained inconclusive within the run's token limit.
Four cases show that the released adapter and prompt can produce useful issue memos across the tested patterns. They are a working diagnostic, not an estimate of accuracy across a full financing-document set. The next meaningful test is a larger held-out set of paired provisions that reflects the deal terms and drafting forms a team uses.
What to build with it
Start with a comparison table. Put the term-sheet language in one column, the draft language in another, and the model's issue memo in a third. Require the reviewer to accept, edit, or dismiss each item.
That review record becomes useful data. It shows which deal terms are easy to compare, which pairings fail, and which proposed corrections survive review. It also keeps the source sections visible instead of turning the model's memo into an unsupported conclusion.
The adapter, recorded examples, prompt format, and offline runner are available under Apache 2.0 in the LiTiL Financing Review 9B repository.