SAFEMARXAI Training
Training

Four stages, and GRPO.

The model learns language first, then the shape of a fair answer, then is rewarded for answers a program can check.

01Pretrain

Read widely

About 3.1 billion tokens of general English, with the fairness library woven in, at most four copies.

02Fine-tune

Learn inclusive answers

Worked examples in inclusive language, each one paired with genders and groups swapped.

03GRPO

Reward fairness

Answers a program scores as fair under the requested standard are made more likely.

04Use

Fairness at work

Once tested, it checks how wages, rent, care and aid are shared, and who is left out.

Stage 1 · Pretraining · built, not to completion

3.17 billion tokens on one Spark

The reading is 3.12 billion tokens of FineWeb-Edu, with the public-domain fairness library woven in between documents, at most four copies of it. Each step reads 64 windows of 2,048 tokens, so the run is about 24,200 steps.

The learning rate warms up over the first 2%, holds steady, then falls to zero over the last 20%. Copies of the model are kept at 1.0B and 2.0B tokens, so a usable early model exists after about a third of the run.

first: a 50-step timing run, then a 300-step check of two learning rates

keep 1.0B keep 2.0B 0 1B 2B 3B tokens read peak 0 decay 2.53B → 3.17B
Stage 2 · Fine-tuning · not built

Teach fair, inclusive answers

Before any reward, the model studies worked examples written in gender-inclusive language. Every problem comes in matched pairs with genders and groups swapped, so the model sees that a fair split never depends on who someone is. Unpaid care work counts as work.

A worked answer has four parts
  1. The split. Who gets what, adding up to the pool.
  2. The standard. Named, for example need or maximin.
  3. The check. Why the numbers meet it, whatever anyone’s gender or group.
  4. The tradeoff. Who would gain or lose under another standard.
Stage 3 · GRPO · not built

Group relative policy optimisation

For each problem the model writes a group of eight answers. A program scores each one. Every answer is then compared with its own group: those above the group's average become more likely, those below become less likely. A penalty keeps the model close to its Stage 2 self, so it cannot drift into nonsense that happens to score.

Gender fairness is trained in here too: every answer is also scored with each citizen’s gender and group swapped, and an answer whose split changes earns nothing. GRPO needs no second model to estimate value, which suits a reward a program can compute. Problems start easy, two people and one standard, and grow as the success rate rises, because a group where every answer scores zero teaches nothing.

mean .53 .92 .55 .10 .81 .47 .00 .88 .52
What the scoring program checks
  1. Every share adds up to the pool. Nothing is skimmed off and no citizen is short-changed.
  2. The standard of justice the question asked for really holds, for every citizen in the numbers.
  3. Nothing is left unused that could lift someone up at no cost to anyone else.
  4. The answer names its standard of justice and says who would gain or lose under another.
  5. Swapping a person's name, gender or group leaves their share unchanged, or the reward is zero.
Stage 4 · Use · after testing

One society, every citizen in harmony

Once its tests are published, the model goes to work for social justice: checking how wages, rent, care and aid are shared, and showing who is left out. Every answer shows its rule, so a society can see its own standards plainly and hold them higher.

Status
  1. Pretraining. Not to completion.
  2. Fine-tuning. Not to completion.
  3. GRPO. Not to completion.