ReedZyd/predictionbench-ot10k-groups

PredictionBench – OpenThoughts-Agent 10K groups (terminal track)

Everything needed to fine-tune Qwen3-4B on the selected 1k-trajectory groups with the OpenThoughts-Agent recipe.

Data (data/)

• train_10k.parquet: the 10,000-row OpenThoughts-Agent-SFT-10K pool (column messages, ChatML-style turns incl. <think> in every assistant turn).
• val.parquet: shared validation split.
• groups_<agent>_ot10k_qwen3-4b_r<seed>/group_{1..5}.parquet: five disjoint 1,000-row groups chosen by the selection agent (group_assignment.csv = row index into train_10k.parquet + group id; method.md = the agent's own description; groups_index.json = indices). groups_random_seed0/ = random baseline (3 groups).
• Ranking convention: group 1 = the agent's predicted BEST post-training result, group 5 = predicted worst.

Recipe (recipe/, frozen — do not change hyper-parameters)

Qwen/Qwen3-4B (hybrid, snapshot 1cfa9a72), full-parameter SFT, lr 4e-5 cosine, warmup 0.1, weight decay 0, adam beta2 0.98, max_grad_norm 1e-4, global batch 96 (= GPUS x PER_DEVICE_BS 1 x GRAD_ACCUM), 7 epochs, cutoff 32768 tokens, assistant-only loss over EVERY assistant turn including <think> (tokenizer dir tokenizer_genmask_qwen3-4b-allthink), no shuffling (SequentialSampler), bf16, FSDP. One 1k group = 77 optimizer steps; ~3 h on 2xH100, ~1.5 h on 4xH100.

Run (any number of GPUs, set GRAD_ACCUM = 96 / NGPUS):

export NGPUS=2 GRAD_ACCUM=48 PER_DEVICE_BS=1 MAX_LEN=32768 LR=4e-5 EPOCHS=7 WARMUP_RATIO=0.1 WEIGHT_DECAY=0 \
       DATA_ROOT=$PWD/data TRAIN_FILE=$PWD/data/groups_<agent>_ot10k_qwen3-4b_r1/group_<k>.parquet VAL_FILE=$PWD/data/val.parquet \
       OUTPUT_DIR=<ckpt_root>/ot10k-<agent>r1-otrecipe-g<k>-4b RUN_NAME=ot10k-<agent>r1-otrecipe-g<k>-4b \
       MODEL=Qwen/Qwen3-4B TOKENIZER_DIR=$PWD/recipe/tokenizer_genmask_qwen3-4b-allthink ACCEL_CONFIG=$PWD/recipe/fsdp.yaml
bash recipe/run_sft.sh $NGPUS

submit/ot-split-flex.sub shows the exact HTCondor environment we use. Python env: torch 2.12 / transformers 5.12 / trl 1.6 / accelerate 1.14 (see recipe/sft.py header for the pinned behaviour). The final directory must contain model.safetensors*, config.json in serving (4.x) format and the stock Qwen3 tokenizer files (sft.py rewrites them at the end) — it is loaded by vLLM 0.11 for evaluation.

What still needs training

See CHECKPOINT_STATUS.txt (done = checkpoint exists on our cluster, missing/partial = please train). Name the output ot10k-<agent>r<seed>-otrecipe-g<k>-4b and upload the final directory (no optimizer states) back to this repo under checkpoints/.