litillabs/litil-financing-review-9b

LiTiL Financing Review 9B

LiTiL Labs · Legal Models

What this model does

LiTiL Financing Review 9B compares a financing term sheet with draft transaction documents. It identifies where the draft differs from the agreed business term, cites the supplied sections, and proposes a concise correction. Start with short, clearly labeled excerpts covering one deal term at a time.

Use it after the relevant provisions have been extracted from the term sheet and draft. In a financing-review workflow, the model can prepare a first-pass issues list for liquidation preference, notice periods, option-pool size, board composition, and similar terms. Its memo can be displayed beside the source language so a deal team can verify the comparison and decide whether to revise the draft.

• Useful for: comparing financing terms and preparing a first-pass issues memo
• Give it: one agreed term and the corresponding draft provision, with section labels
• It returns: the difference, the source sections, and a proposed correction

At a glance	Detail
Release repository	litillabs/litil-financing-review-9b
Base	Qwen3.5-9B, text architecture
Input → output	Term-sheet and draft excerpts → Markdown issues memo
Format	PEFT LoRA adapter
Demonstrated runtime	Apple MPS · BF16 · Transformers 5.9.0 · PEFT 0.19.1

Example · Quickstart · Training · Evaluation · Use notes

Input format and output example

Recorded synthetic comparison, September 10, 2026. The user requests a memo of at most 100 words, based only on these excerpts:

Compare the following provisions and return a memo of at most 100 words.

Term sheet §1: Series Seed has a 1x non-participating liquidation preference.

Draft charter §2: Series Seed first receives 2x its purchase price and then
participates pro rata with common in remaining proceeds.

The adapter returned:

- Liquidation preference: Term sheet §1 provides a 1x non-participating
  preference; draft charter §2 provides a 2x participating preference.
  Revise draft §2 to 1x non-participating to match the term sheet.

The full prompt is in examples.json; recorded_examples.json contains the observed response and three additional comparisons. Displayed responses end at the next role boundary and omit special tokens.

Quickstart

Read recorded_examples.json to inspect the outputs without model files. From the prepared inference repository, its runner can display the selected record or generate a new response:

python -m pip install -r requirements.txt
python demo.py --show-recorded --example ecvc_liquidation
python demo.py --base /models/Qwen3.5-9B \
  --adapter /models/litil-financing-review-9b --example ecvc_liquidation

The second command requires separately available base weights, adapter weights and tokenizer. The documentation ZIP supplies information rather than model files. --list-examples lists cases; --input accepts an object matching examples.json. Tested dependencies are recorded in requirements.txt.

Training

Supervised LoRA tuning used 88 synthetic financing examples and 25 optimizer steps on Apple Silicon. Rank is 16, alpha 32 and dropout 0.05, targeting attention and MLP projections. The saved training configuration sets a 512-token sequence length.

Training messages were reproduced from the authored synthetic generator. The reconstruction accounts for all 88 post-training rows, and the generator reads no private source input. No private client or user data appears in the reviewed post-training set. This M5 adapter is a separate run from the Dense and FullGold financing variants. Its prompt uses literal <|system|>, <|user|> and <|assistant|> role strings, retained by the runner.

Evaluation

A fresh four-case diagnostic used the original full base on Apple MPS with greedy generation. The first assistant responses correctly compared liquidation preference, notice periods, option-pool size and a matching board provision.

Diagnostic	Observed result
Four supplied comparisons	Correct core comparison in 4/4 first responses
Focused liquidation follow-up	Complete correction in 57 generated tokens
Matched-base comparison	Final-answer quality comparison remains inconclusive

The follow-up used a 1,024-token limit and role-boundary stops for both arms. The base had not produced its final memo within that limit. These short examples demonstrate a usable response style; they do not estimate accuracy across financing documents.

Runtime and version

The release is a PEFT LoRA adapter and requires the Qwen3.5-9B base model. The selected adapter is approximately 116 MB. The retained package was tested with Apple MPS in BF16 using Transformers 5.9.0 and PEFT 0.19.1.

• Base repository: Qwen/Qwen3.5-9B
• Tested base snapshot: c202236235762e1c871ad0ccb60c8ee5ba337b9a
• Adapter SHA-256: 8c16a9fb79f31499cf9d4eb05a120048f20aa3d77e80f54ccac76ee84d517cca
• Card version: September 11, 2026
• Organization: LiTiL Labs

Use notes

• Supply both provisions with section labels and ask for a bounded comparison. The demonstrated scope is short excerpts, not an entire closing set.
• Preserve the literal training prompt and stop at the next role boundary to return one memo.
• Verify the cited language and any proposed correction against the supplied documents; the model should not fill in missing deal terms.