Updated readme

Fix eval decode prompt length slicing

Fix GRPO reward mapping and evaluation generation

Include prompt column for GRPOTrainer

Use finite datasets.Dataset for GRPOTrainer compatibility

Yield tensors from GRPO dataset generator

Fix GRPO dataloader batching on Kaggle

Fix GRPO sample id mapping and Kaggle training setup

Fix division by zero in final avg improvement calculation

Fix category enum: use lowercase for ProbeAction categories

Fix eval_report.py: use step() return value and extract reward from observation

Remove unsupported data_collator argument from GRPOTrainer

Add data collator to GRPOTrainer for on-the-fly prompt tokenization

Revert to raw prompt format—let GRPOTrainer handle tokenization

Convert tokenized outputs to torch tensors for proper batching

Tokenize prompts in dataset generator for GRPOTrainer compatibility

Fix prompt format: convert from chat list to string for GRPOTrainer compatibility

Fix dataset format: only yield 'prompt' field to avoid tensor concatenation errors

Remove tokenizer argument from GRPOTrainer—not supported in installed TRL version

Fix path bootstrap in train_grpo.py: use parent.parent to reach project root

Fix GRPO batch size config for Kaggle P100: batch_size=4, grad_accum=2 (global=8, divisible by num_generations=2)

Add eval_report.py for before/after training comparison

Fix JSON parsing and environment bugs

Code improvenets

refactor: remove legacy architecture, promote clean structure to repo root