Downloads · 30 days
0
abdulmuminyqn/drafter-finance-critic-1.7b-mlx-lora
drafter-finance-critic-1.7b-mlx-lora is a text generation model from abdulmuminyqn. Use it when you need the model to write or continue text. It is set up for mlx. The card lists the license as apache-2.0.
This is a narrow claim-extraction adapter for Qwen/Qwen3-1.7B-MLX-4bit. Given a client-facing finance draft and its source ledger, it emits structured claims for a deterministic verifier. It is not a general finance m…
Downloads · 30 days
0
Access
Public
Updated Jul 16, 2026
Repo size
19.9 MB
Likes
0
Public
Click a slice to open those files.
.safetensors19.9 MB · 100%
From the Hugging Face model README
This is a narrow claim-extraction adapter for Qwen/Qwen3-1.7B-MLX-4bit.
Given a client-facing finance draft and its source ledger, it emits structured
claims for a deterministic verifier. It is not a general finance model, a
financial adviser, or an autonomous decision-maker.
The adapter was trained entirely on deterministic synthetic data. No customer, employer, interview, or confidential company data is included.
Use this adapter as one component in a guarded workflow:
The expected output is a JSON object with a claims array. The companion app
uses ordinary text generation rather than provider-enforced JSON mode and accepts
JSON wrapped in prose or Markdown fences.
Apple Silicon and mlx-lm==0.31.3 are the tested runtime. Download the public
adapter, then load it beside the pinned base model:
hf download abdulmuminyqn/drafter-finance-critic-1.7b-mlx-lora \
--local-dir ./drafter-finance-critic-1.7b-mlx-lora
from mlx_lm import generate, load
model, tokenizer = load(
"Qwen/Qwen3-1.7B-MLX-4bit",
adapter_path="./drafter-finance-critic-1.7b-mlx-lora",
)
For the companion app, clone its experiment branch and place the adapter at the documented local path:
git clone --branch experiment/finance-critic-lora \
https://github.com/Abdulmumin1/drafter-agent.git
cd drafter-agent
uv sync --extra train
hf download abdulmuminyqn/drafter-finance-critic-1.7b-mlx-lora \
--local-dir training/adapters/critic-qwen3-1.7b-v2
Start an OpenAI-compatible local server without a global adapter:
uv run --extra train mlx_lm.server \
--model Qwen/Qwen3-1.7B-MLX-4bit \
--host 127.0.0.1 \
--port 8080 \
--max-tokens 900 \
--chat-template-args '{"enable_thinking":false}'
export OPENAPI_CRITIC_BASE_URL="http://127.0.0.1:8080/v1"
export OPENAPI_CRITIC_MODEL="Qwen/Qwen3-1.7B-MLX-4bit"
export OPENAPI_CRITIC_ADAPTER="training/adapters/critic-qwen3-1.7b-v2"
uv run drafter critic-assess
The app sends the adapter on critic requests only, leaving drafting/revision free to use another OpenAI-compatible model.
Qwen/Qwen3-1.7B-MLX-4bit21457c6f51ed54a7c16e988c0844db973815c13748f49ed1e122ebc8ee7a0e916c4bec4e7baadd73135fbbdf212629ec27135cb7The exact config and dataset manifests are included in this repository. Raw JSONL is regenerated from the open source curriculum builder and is not bundled here.
Results below use greedy decoding and four held-out examples per family. Checkpoint selection used validation loss before either test rematch.
| Metric | Earlier 100-step adapter | This adapter |
|---|---|---|
| Valid JSON | 100.0% | 100.0% |
| Exact claim-span F1 | 74.6% | 73.5% |
| Full-claim accuracy | 33.8% | 47.1% |
| Unsupported-claim recall | 75.0% | 83.3% |
| Offer-only accuracy | 0.0% | 100.0% |
| Invented citation IDs | 0 | 0 |
| Metric | Earlier 100-step adapter | This adapter |
|---|---|---|
| Valid JSON | 100.0% | 100.0% |
| Exact claim-span F1 | 61.7% | 76.4% |
| Full-claim accuracy | 22.7% | 38.6% |
| Unsupported-claim recall | 75.0% | 95.8% |
| Offer-only accuracy | 0.0% | 100.0% |
| Invented citation IDs | 0 | 0 |
In the companion app, the adapter plus its conservative optional-field parser and deterministic verifier caught 7/7 adversarial corruptions with 7/7 safe actions and zero critic failures. That is a system result, not a standalone model score.
The measured tradeoff is explicit: exact claim-span F1 on the original benchmark fell 1.1 percentage points while full-field accuracy, unsupported-claim recall, and offer handling improved. Deterministic verification and human review remain required.