FinSentQwen-14B

Two sizes. This 14B is the stronger model on the four aspects, by 0.021 (within noise). On overall sentiment the two sizes are level. FinSentQwen-4B costs about a third as much to run.

FinSentQwen-14B reads an English news passage and a named company, and labels how the passage bears on that company. Each label is positive, negative or no signal, and there are five of them, all from one forward pass:

Aspect	What it covers
financial performance	revenue, profit, losses, results
demand and commercial traction	customers, orders, sales, market share
operations, supply, and capacity	production, supply chain, outages, expansion
outlook and expectations	guidance, forecasts, expected developments
overall company sentiment	the passage's net stance on the company

This repository holds a LoRA adapter plus a 3-class score head for Qwen/Qwen3-14B (revision 40c069824f4251a91eefaf281ebe4c544efd3e18), predict.py, and thresholds.json (labels and input settings). It was trained only on open-licence news, with labels from GPT-6.1. Both FinSentQwen sizes use the same data and recipe.

Results

Scored on a sealed test of 400 news passages (2022–2024; VOA, Global Voices, 360info, Factly, Freedom of the Press Foundation) that was never used for training. The answer key is GPT-6.1, the same labeller and prompt as the training data, so the score shows how well the model learned its teacher (26 passages GPT-6.1 declined count as "no signal"). Claude Opus also labelled the test, twice and double-blind, as an independent check. Macro-F1 over the three classes, with labels by majority vote (see Decision rule):

Task	FinSentQwen-14B	FinSentQwen-4B	Claude Opus labels
Four aspects (1,600 cells)	0.837	0.816	0.851
Overall company sentiment (400)	0.806	0.810	0.803

Against the Opus labels instead:

Task	FinSentQwen-14B	FinSentQwen-4B	GPT-6.1 labels	Opus ceiling
Four aspects	0.829	0.818	0.851	0.960
Overall company sentiment	0.773	0.769	0.803	0.936

The Opus ceiling is the second Opus labelling scored against the first. Opus calls "no signal" on overall sentiment more often than GPT-6.1 does, which is most of the drop on that row.

• 14B vs 4B (GPT-6.1 key): the 14B is ahead by 0.021 on aspects (paired bootstrap over passages, 95% CI −0.015 to +0.059) and behind by 0.004 on overall sentiment (CI −0.033 to +0.024). That is one training seed each.
• 88% of aspect cells in the key are "no signal". The model calls 178 cells positive or negative, against 185 in the key. Per-class F1: positive 0.76, negative 0.78, no signal 0.97.
• These scores are from a later look. See How the rule was chosen below.

By source (GPT-6.1 key):

Slice	Aspects	Overall
VOA (251 passages)	0.862	0.839
Common Pile publishers (132)	0.776	0.746
2022 top-up (17)	0.897	0.800

The 2022 slice is too small to read.

Calibration (raw softmax against the GPT-6.1 key, 10 equal-mass bins): classwise ECE 0.007 on aspects and 0.048 on overall sentiment.

Cost: fit 4.3 h on 2×T4. Inference is about 843 T4-seconds per 1,000 passages (≈ $0.08 at $0.35/T4-hour).

Use

pip install torch transformers peft bitsandbytes accelerate
python predict.py input.jsonl > output.jsonl

Each input line is {"id": ..., "company": "...", "text": "..."}. Each output line adds every aspect's label and its probs (positive, negative, no signal). The base model loads in 4-bit NF4, as it was trained, on one CUDA GPU with about 12 GB. predict.py reproduces the scored test predictions from this adapter.

Input format. The model sees Company: {company}\n\nPassage:\n{text}, then \n\nAspect: {aspect}\nSignal: for each of the five aspects. The head reads the last token of each aspect line. Input is capped at 1,024 tokens, and only the passage is cut.

Decision rule. Each cell gets the class with the largest softmax share, a plain majority vote. Nothing is tuned or corrected, and probs is the raw softmax. thresholds.json holds the classes, aspect strings and input settings.

How the rule was chosen. The test was first scored once against the Opus labels, with a "no signal" threshold tuned on a news dev set and frozen beforehand: t 0.09 for aspects and 0.15 for overall, scoring 0.734 / 0.835. That threshold did not transfer to the test. On 2026-11-05 we switched to the plain vote, because it has nothing to tune, and made GPT-6.1 the answer key, so the model is graded by the labeller it was trained on. A prior correction toward Opus's label mix was tried in between and dropped for simplicity: it scores 0.838 / 0.811 against Opus but 0.817 / 0.810 against GPT-6.1. All of this came after the first result was known, so read the scores as less clean than a fresh test. A new sealed test is planned.

Training

• Data: 5,134 English news passages, each with a named company, and 25,670 labels (5 per passage) from GPT-6.1 (gpt-6.1-sol via Codex). The labels and the text are in neoyipeng/FinSentQwen-data. No SEC filings or press-release exhibits were used anywhere.
• Recipe: QLoRA (NF4, r 16, alpha 32, dropout 0.05, all attention and MLP projections) plus a 3-class score head; lr 1e-4, linear warmup, 2 epochs, effective batch 16 passages, fp16, 2×T4.
• Choosing size and epochs: a scaling grid (Qwen3 0.6B / 4B / 14B × three data sizes × 3 seeds), scored on a 1,000-passage dev holdout, picked 14B at 2 epochs under a rule fixed before the grid ran. 14B led 4B by 0.013 on aspects and 0.007 on overall sentiment there. A second epoch helped; a third did not.
• One seed is published, and it is the one that was scored.

Sources and licences

The training and test text comes from these sources. Passages are excerpted and normalized.

Source	Licence	Link
Voice of America	Public domain (VOA-produced)	https://www.voanews.com/p/5338.html
Multilingual Open Text v1.11 (BLT Lab), VOA collection	CC BY 4.0	https://github.com/bltlab/mot
Global Voices (via Common Pile)	CC BY 3.0	https://globalvoices.org
360info (via Common Pile)	CC BY 4.0	https://360info.org
Factly (via Common Pile)	CC BY 4.0	https://factly.in
Freedom of the Press Foundation (via Common Pile)	CC BY 4.0	https://freedom.press
Balkan Diskurs (via Common Pile)	CC BY 3.0	https://www.balkandiskurs.com

Per-passage attribution (author, title, URL, licence) is in the dataset. The adapter is released under Apache-2.0, as is the Qwen3 base model.

Limitations

• English news only, with a company named for the passage. It is not tested on filings, social media, transcripts or other languages.
• The labels are silver: GPT-6.1 for training and the main test key, Claude Opus as a check. The two agree at 0.851 / 0.803 on the test.
• It doesn't read prices or market reaction, and it isn't investment advice.