FilingFigures Qwen3.5-4B (NVFP4)
Qwen3.5-4B fine-tuned to read a 10-K and return six financial figures as JSON, quantized to 4-bit NVFP4 for vLLM. On held-out companies it gets 98.8% of figures right, ties Qwen3.5-397B and costs $0.07 per 1,000 filings on one RTX 5090.
Code, prompt, data and full results: github.com/nathankyryk/filing-figures. The LoRA adapter is at kyryk/filing-figures-qwen3.5-4b-lora.
Results
Test split of kyryk/filing-figures: 207 filings from 36 companies not seen in training, 1,186 figures, run once.
| Model | Figures right | Filings with all 6 right | $ per 1,000 filings |
|---|---|---|---|
| Qwen3.5-4B, no fine-tuning | 95.4% | 83% | $0.11 |
| This model (fine-tuned, NVFP4) | 98.8% | 93% | $0.07 |
| Qwen3.5-397B | 98.7% | 93% | $3.66 |
| Claude Opus 5.5 | 100.0% | 100% | $29.86 |
A figure is right when it's within 0.5% (EPS: $0.005) and its period end date is exact.
Use it
Needs an NVIDIA Blackwell GPU (RTX 50-series, B200) for NVFP4. Tested with vLLM 0.32.0 on an RTX 5090.
vllm serve kyryk/filing-figures-qwen3.5-4b-nvfp4 --served-model-name filing-figures \
--max-model-len 12288 --language-model-only \
--default-chat-template-kwargs '{"enable_thinking": false}'
Then send it a 10-K with the repo's CLI. This fetches the filing from EDGAR, builds the input and scores the answer against XBRL:
git clone https://github.com/nathankyryk/filing-figures && cd filing-figures && uv sync
uv run figures extract --ticker SBUX --year 2025 --endpoint http://localhost:8000/v1 --model filing-figures
The model expects the exact prompt it was trained with: the system message in src/filing_figures/prompts.py, then the filing as built by src/filing_figures/inputs.py. It answers with one JSON object:
{"revenue": {"value": 36176200000, "scale": "millions", "period_end": "2024-09-29", "fiscal_year": 2024}, "net_income": {...}, "eps_diluted": {...}, "total_assets": {...}, "shares_outstanding": {...}, "operating_cash_flow": {...}}
How it was made
- Fine-tuning: bf16 LoRA (rank 16, Ξ± 32, lr 2e-4) with Unsloth on 940 training filings. The loss covers the answer only. The checkpoint after half an epoch was kept: every checkpoint scored 100% on validation, so the earliest won.
- Quantization: llm-compressor NVFP4, calibrated on 256 training filings.
lm_head, the vision tower and Qwen3.5's linear-attention layers stay in bf16. Validation accuracy: 99.8% vs 100% in bf16. - Speed: 4.6 filings/s at 32 in flight on one RTX 5090 (~5,000 tokens in and ~300 out per filing).
Limitations
- Trained and tested on non-financial S&P 500 10-Ks from 2019β2024. Banks and insurers weren't included.
- Its remaining test errors are mostly definitional: "Sales and other operating revenues" vs total revenues, and net income before vs after preferred dividends.
- Answers only from the text it's given. It wasn't tested on PDFs, scanned filings or other report types.
- Downloads last month
- 280