AdaptiveSFT: baseline SFT-loss runs

Full fine-tuning runs of seven token-level SFT objectives (and the untrained base model) swept over learning rates, with their checkpoints, training logs, sampled eval outputs and scores, plus the exact training and eval sets. Code: https://github.com/jiarui-liu/adaptive-reasoning-loss (framework arl/, the commit of each run is in the tables).

Layout

data/sft/v1/<set>.jsonl                 training sets ({prompt, response[, source]})
data/sft/v1/eval_sets/<bench>.jsonl     eval sets ({problem, answer, kind, ...})
runs/<grid>/<method>_lr<lr>/
    final/                checkpoint (safetensors + tokenizer), loads with transformers / vLLM
    args.json             all training arguments
    train_log.jsonl       one row per optimizer step: loss, nll, grad norm, lr, tokens, method statistics
    train.log, eval.log   raw stdout of training / evaluation
    train.sbatch, eval.sbatch   the exact slurm scripts that produced the run
    eval/<bench>.jsonl    every sample per eval problem, finish reasons, token counts
    eval/scores.json      pass@1 / pass@k, mean length, truncation rate per bench

base rows evaluate the untrained starting model (no final/). Data builders: experiments/data/ (build_sets.py for the math sets, build_tulu.py for the Tulu subset and IFEval / GSM8K / MMLU / HumanEval+). Prompt formats: math sets use Problem: {prompt}\n\nSolution:\n; tulu3_50k uses <|user|>\n{prompt}\n<|assistant|>\n (arl/data.py).

Methods

method	per-token objective
nll	standard cross-entropy
dft	NLL weighted by the model's own (stop-gradient) probability of the target token
focal	focal loss, down-weights tokens the model is already confident on
trimsft	NLL with part of the tokens trimmed from the loss
prift	weights from the frozen starting model's probabilities (mode mass)
asft	DFT weighting plus a KL anchor to the frozen starting model
psft	PPO-style clipped ratio against the model a few steps ago (reference refreshed every 4 steps)

Scores are pass@1 in % (mean over samples). Runs appear here once their evaluation has finished.

s1_numina_q17

math reasoning distillation: Qwen3-1.7B-Base on numina_cot_20k (20k NuminaMath-CoT problems, one CoT solution each), 3 epochs, 64 seqs/step, max 4096 tokens; eval on math benchmarks

run	method	model	lr	method args	aime24	aime25	amc23	math500	minerva	olympiadbench	code commit
nll_lr5e-6	nll	Qwen3-1.7B-Base	5e-06	{}	1.7	1.9	26.6	52.5	17.9	22.6	28b0306 (inferred from start time)
nll_lr1e-5	nll	Qwen3-1.7B-Base	1e-05	{}	1.7	1.0	27.0	52.8	17.6	22.4	28b0306 (inferred from start time)
nll_lr2e-5	nll	Qwen3-1.7B-Base	2e-05	{}	1.0	2.1	26.9	50.0	16.5	21.7	28b0306 (inferred from start time)
nll_lr5e-5	nll	Qwen3-1.7B-Base	5e-05	{}	1.0	0.8	22.8	47.0	16.2	17.8	28b0306 (inferred from start time)

Sources and licenses

Base model: Qwen/Qwen3-1.7B-Base (Apache-2.0). Training data are subsets of AI-MO/NuminaMath-CoT, open-r1/OpenR1-Math-220k and allenai/tulu-3-sft-mixture (ODC-BY; sub-sources keep their own licenses). Eval sets: MATH-500, Minerva Math, OlympiadBench, AMC23, AIME24/25, IFEval, GSM8K, MMLU, HumanEval+.