This is the final loadable model root from the 400-step SDFT run qwen3-4b-instruct-2507-sdft-mctsstyle-parity400-concisehint-lr1em6-cumem0-20260615. It starts from Qwen/Qwen3-4B-Instruct-2507 and trains on LorMolf/SPSD-RL using TRL experimental self-distillation with MCTS-style privileged contexts.

Training Summary

Method: SDFT over prompt plus privileged MCTS-context rows.
Budget: 400 optimizer steps with generation_batch_size=128, or 51,200 generated completions.
Learning rate: 1e-6.
Runtime: four H200 GPUs, DDP, NCCL_CUMEM_ENABLE=0, USE_FSDP=false.
Student concise hint: enabled.
generate_from_teacher: false.
Train-time eval: disabled (eval_strategy=no) because the TRL SDFT eval path is incompatible with raw prompt/privileged-context batches in this setup.
W&B run: x2tz41er in project olmo-spiral-sdft.

The final trainer summary recorded train loss 0.05229, runtime about 1.133e+05 seconds, train samples/s 0.452, and train steps/s 0.004.

Artifact Scope

This Hugging Face repository intentionally contains only the final loadable model payload: model weights, tokenizer/config files, generation config, chat template, and this model card. Intermediate checkpoint-* directories, optimizer/scheduler state, trainer state, logs, caches, and dataset materialization artifacts are not included.

Evaluation

Post-hoc reasoning and OpenReward/CRPT game evaluations should use the repository standard boxed SFT/SDFT profile (sft_boxed_forced) so visible reasoning is preserved while final answers are canonicalized to Final answer: \boxed{...} when needed. No post-hoc benchmark results are included in this model card at upload time.

LorMolf/SPSD-RL-Qwen3-4B-Instruct-2507-SDFT-MCTSStyle-ConciseHint

SPSD-RL Qwen3-4B Instruct 2507 SDFT MCTSStyle ConciseHint

Training Summary

Artifact Scope

Evaluation

LorMolf/
SPSD-RL-Qwen3-4B-Instruct-2507-SDFT-MCTSStyle-ConciseHint