update: results table, 0.5B model links, citation year 2026

docs: detailed README with curriculum, reward table, results, usage

rollback: revert to last working Dockerfile and train.py

fix: proper GRPO with trl 0.12.2 no-deps + force hub downgrade

fix: custom training loop without TRL dependency

fix: force reinstall huggingface_hub 0.24.7 after deceit_env

fix: pin huggingface_hub 0.24.7, install trl with --no-deps

fix: trl 0.12.2 has GRPOTrainer, pin all deps before trl install

fix: try multiple import paths for GRPOConfig

fix: install transformers 4.46.0 BEFORE trl so trl doesnt upgrade it

fix: bust docker cache force reinstall trl 0.11.4

fix: trl 0.11.4 + transformers 4.46.0 + processing_class

fix: trl 0.9.4 + transformers 4.41.2 compatible versions

fix: add torch version check in Dockerfile

fix: remove tokenizer arg from GRPOTrainer

fix: tokenizer not processing_class, torch cu121 for GPU

fix: correct trl version with GRPOConfig

fix: trl 0.13.0, remove verify steps

fix: cu124 not cu118 for A100 CUDA 12.9 driver

fix: trl 0.12.2 + torch 2.4.0

fix: CPU fallback when no GPU detected

fix: trl 0.15.0 definitely has GRPOConfig

fix: trl 0.9.6 + bitsandbytes 0.43.1 cu118

fix: trl 0.8.6 has GRPOConfig, compatible with torch 2.1.2

fix: remove multiline python heredoc from Dockerfile

fix: find deceit_env package location and copy data correctly

fix: back to python:3.10-slim for GPU, fix deceit_env path

fix: deceit_env module path and PYTHONPATH

fix: nvidia cuda base with python3.10 installed

fix: use huggingface transformers-pytorch-gpu base image

fix: revert to torch 2.1.0 cu121 with trl 0.7.4 - versions that worked before

fix: simplify dockerfile no version pinning

fix: accelerate 0.34.2 exists, 0.35.0 does not

fix: trl 0.12.0 has GRPOTrainer, compatible with torch 2.4.0

fix: pin trl 0.11.0 compatible with torch 2.4.0

fix: upgrade torch to 2.4.0 with CUDA 12.4 support

fix: run train.py instead of evaluate.py

improve: abstention penalty, better prompt, mixed curriculum, more steps

evaluate: switch to 0.5B model comparison, 200 episodes

fix: set N_EPISODES=200 constant (was still 30)

update: increase evaluation to 200 episodes per model

fix: parse_action confidence bug, numeric answers bug, missing reasoning field bug

fix: debug model output parsing in evaluation

add: evaluate 1.5B base vs trained, upload chart to HF Hub

update: 500 steps L1 + 300 steps L2, higher lr for 1.5B

fix: add matplotlib, split COPY into separate lines

fix: copy evaluate.py into Docker image

add: evaluate 1.5B base vs trained, upload charts

fix: remove misplaced import inside GRPOConfig args

fix: auto-detect bf16 support