wehub/microsoft--unilm
1
0
Fork 0
Code Issues Pull Requests Actions Deployments Models Agent Notes Packages Projects Releases Wiki Activity

chore: import upstream snapshot with attribution

Browse Source
This commit is contained in:
wehub-resource-sync
2026-07-13 13:24:13 +08:00
commit 1037506f2e
6050 changed files with 1731598 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
PFPO/scripts/math_scale/mscale/4o_pipeline.sh
+15
View File
@@ -0,0 +1,15 @@
python scripts/math_scale/construct_prefer_pair.py \
--input_file "../msranlpintern/reward_modeling/experiments/mathstral.mathscale4o.process-dpo.iter0.V100.tp8dp48.v2.2.fix.s42/checkpoint-600/mathscale4o/mscale-v0.1-300k/mscale.v0.1.300k.v1.0.n10.tem1.0.p1.0.*-of-512.s42.json" \
--output_file "../msranlpintern/reward_modeling/experiments/mathstral.mathscale4o.process-dpo.iter0.V100.tp8dp48.v2.2.fix.s42/checkpoint-600/mathscale4o/mscale-v0.1-300k/mscale.v0.1.300k.v1.0.n10.tem1.0.p1.0.prefer_pair.4o.json"
#Total number of items: 298275
#Acc: 0.7078333752409689 Pass at k: 0.8356214902355209
#No positive solutions: 49030 / 298275 No negative solutions: 148203 / 298275
#Num pairs: 1650384
python scripts/math_scale/construct_process_rm_sample_gd.py \
--input_file "../msranlpintern/reward_modeling/experiments/mathstral.mathscale4o.process-dpo.iter0.V100.tp8dp48.v2.2.fix.s42/checkpoint-600/mathscale4o/mscale-v0.1-300k/split-1024/mscale.v0.1.300k.v1.0.n10.tem1.0.p1.0.upper0.7.r0.3.sample10.filter_same.prefix_completion.n3.tem1.0.p0.9.*-of-1024.json" \
--output_file ../msranlpintern/reward_modeling/experiments/mathstral.mathscale4o.process-dpo.iter0.V100.tp8dp48.v2.2.fix.s42/checkpoint-600/mathscale4o/mscale-v0.1-300k/mscale.v0.1.300k.v1.0.n10.tem1.0.p1.0.upper0.7.r0.3.sample10.filter_same.process_rm.4o.json \
--num_workers 24
#Counter({0: 648447, 3: 465424, 2: 184121, 1: 178847, 6: 3645, 4: 1485})
#Missing 308 items in the response data.
#Counter({0: 648147, 3: 465044, 2: 183998, 1: 178757, 6: 3642, 4: 1484})
PFPO/scripts/math_scale/mscale/construct_prm_sc_pair.sh
+39
View File
@@ -0,0 +1,39 @@
export OUTPUT_PREFIX_PATH=/mnt/fangkai_blob/reward_modeling/
p=$1
echo "Constructing process_rm sample for split 0"
echo "p" $p
#python scripts/math_scale/construct_process_rm_sample_sc.py \
# --input_file "${OUTPUT_PREFIX_PATH}/experiments/mathstral.mathscale4o.process-dpo.iter0.V100.tp8dp48.v2.2.fix.s42/checkpoint-600/mathscale4o/mscale-v0.1-300k/split-1024/mscale.v0.1.300k.v1.0.n10.tem1.0.p1.0.upper0.7.r0.3.sample10.filter_same.prefix_completion.n3.tem1.0.p0.9.*-of-1024.json" \
# --output_file ${OUTPUT_PREFIX_PATH}/experiments/mathstral.mathscale4o.process-dpo.iter0.V100.tp8dp48.v2.2.fix.s42/checkpoint-600/mathscale4o/mscale-v0.1-300k/mscale.v0.1.300k.v1.0.n10.tem1.0.p1.0.upper0.7.r0.3.sample10.filter_same.process_rm.sc-p$p.azure.json \
# --response_file_for_sc "${OUTPUT_PREFIX_PATH}/experiments/mathstral.mathscale4o.process-dpo.iter0.V100.tp8dp48.v2.2.fix.s42/checkpoint-600/mathscale4o/mscale-v0.1-300k/mscale.v0.1.300k.v1.0.n10.tem1.0.p1.0.*-of-512.s42.json" --response_id_field id --num_workers 128 --top_p $p#
#python scripts/math_scale/construct_process_rm_sample_sc.py \
# --input_file "${OUTPUT_PREFIX_PATH}/experiments/mathstral.mscale-v0.1-300k.process-dpo-sc.p0.5.iter1.V100.tp4dp32.v3.1.s42/checkpoint-500/mathscale4o/mscale-v0.1-300k/split-1024/mscale.v0.1.300k.v1.0.n10.tem1.0.p1.0.upper0.7.r0.3.sample10.filter_same.prefix_completion.n3.tem1.0.p0.9.*-of-1024.json" \
# --output_file ${OUTPUT_PREFIX_PATH}/experiments/mathstral.mscale-v0.1-300k.process-dpo-sc.p0.5.iter1.V100.tp4dp32.v3.1.s42/checkpoint-500/mathscale4o/mscale-v0.1-300k/mscale.v0.1.300k.v1.0.n10.tem1.0.p1.0.upper0.7.r0.3.sample10.filter_same.process_rm.sc-p$p.azure.json \
# --response_file_for_sc "${OUTPUT_PREFIX_PATH}/experiments/mathstral.mscale-v0.1-300k.process-dpo-sc.p0.5.iter1.V100.tp4dp32.v3.1.s42/checkpoint-500/mathscale4o/mscale-v0.1-300k/mscale.v0.1.300k.v1.0.n10.tem1.0.p1.0.*-of-512.s42.json" --response_id_field id --num_workers 128 --top_p $p
#p=0.5
#Counter({3: 493281, 2: 165717, 1: 102586, 0: 86441})
#Missing 298249 items in the response data.
#Counter({3: 487316, 2: 163641, 1: 101298, 0: 85343})
#p=0
#Counter({3: 522458, 0: 241172, 2: 209424, 1: 172472})
#Missing 748 items in the response data.
#Counter({3: 516169, 0: 238263, 2: 206891, 1: 170353})
#python scripts/math_scale/construct_process_rm_sample_sc.py \
# --input_file "${OUTPUT_PREFIX_PATH}/experiments/mathstral.mscale-v0.1-300k.process-dpo-sc.p0.0.iter2.H100.dp16.v1.3.s42/checkpoint-500/mathscale4o/mscale-v0.1-300k/split-1024/mscale.v0.1.300k.v1.0.n10.tem1.0.p1.0.upper0.7.r0.3.sample10.filter_same.prefix_completion.n3.tem1.0.p0.9.*-of-1024.json" \
# --output_file ${OUTPUT_PREFIX_PATH}/experiments/mathstral.mscale-v0.1-300k.process-dpo-sc.p0.0.iter2.H100.dp16.v1.3.s42/checkpoint-500/mathscale4o/mscale-v0.1-300k/mscale.v0.1.300k.v1.0.n10.tem1.0.p1.0.upper0.7.r0.3.sample10.filter_same.process_rm.sc-p$p.azure.json \
# --response_file_for_sc "${OUTPUT_PREFIX_PATH}/experiments/mathstral.mscale-v0.1-300k.process-dpo-sc.p0.0.iter2.H100.dp16.v1.3.s42/checkpoint-500/mathscale4o/mscale-v0.1-300k/mscale.v0.1.300k.v1.0.n10.tem1.0.p1.0.*-of-512.s42.json" --response_id_field id --num_workers 128 --top_p $p
#Counter({3: 470992, 0: 195492, 2: 178005, 1: 144163})
#Missing 391 items in the response data.
#Counter({3: 470540, 0: 195452, 2: 177903, 1: 144094})
python scripts/math_scale/construct_process_rm_sample_sc.py \
--input_file "${OUTPUT_PREFIX_PATH}/experiments/mathstral.mscale-v0.1-300k.process-dpo-sc.p0.0.iter2.H100.dp16.v1.3.s42/checkpoint-500/mathscale4o/mscale-v0.1-300k/split-1024/mscale.v0.1.300k.v1.0.n10.tem1.0.p1.0.upper0.7.r0.3.sample10.filter_same.prefix_completion.n3.tem1.0.p0.9.*-of-1024.json" \
--output_file ${OUTPUT_PREFIX_PATH}/experiments/mathstral.mscale-v0.1-300k.process-dpo-sc.p0.0.iter2.H100.dp16.v1.3.s42/checkpoint-500/mathscale4o/mscale-v0.1-300k/mscale.v0.1.300k.v1.0.n10.tem1.0.p1.0.upper0.7.r0.3.sample10.filter_same.process_rm.sc-p$p.azure.json \
--response_file_for_sc "${OUTPUT_PREFIX_PATH}/experiments/mathstral.mscale-v0.1-300k.process-dpo-sc.p0.0.iter2.H100.dp16.v1.3.s42/checkpoint-500/mathscale4o/mscale-v0.1-300k/mscale.v0.1.300k.v1.0.n8.tem1.0.p1.0.*-of-512.s*.json" --response_id_field id --num_workers 128 --top_p $p
PFPO/scripts/math_scale/mscale/pipeline.sh
+78
View File
@@ -0,0 +1,78 @@
python scripts/math/deepseek_math_sample_steps.py \
--input_file "../msranlpintern/reward_modeling/experiments/mathstral.mathscale4o.process-dpo.iter0.V100.tp8dp48.v2.2.fix.s42/checkpoint-600/mathscale4o/mscale-v0.1-300k/mscale.v0.1.300k.v1.0.n10.tem1.0.p1.0.*-of-512.s42.json" \
--output_file ../msranlpintern/reward_modeling/experiments/mathstral.mathscale4o.process-dpo.iter0.V100.tp8dp48.v2.2.fix.s42/checkpoint-600/mathscale4o/mscale-v0.1-300k/mscale.v0.1.300k.v1.0.n10.tem1.0.p1.0.upper0.7.r0.3.sample10.filter_same.json \
--upper_step_ratio 0.7 --sample_ratio 0.3 --filter_all_same --sample_over_p 10
# Vanilla outcome DPO
python scripts/math_scale/construct_prefer_pair_sc.py \
--input_file "${OUTPUT_PREFIX_PATH}/experiments/mathstral.mathscale4o.process-dpo.iter0.V100.tp8dp48.v2.2.fix.s42/checkpoint-600/mathscale4o/mscale-v0.1-300k/mscale.v0.1.300k.v1.0.n10.tem1.0.p1.0.*-of-512.s42.json" \
--output_file ${OUTPUT_PREFIX_PATH}/experiments/mathstral.mathscale4o.process-dpo.iter0.V100.tp8dp48.v2.2.fix.s42/checkpoint-600/mathscale4o/mscale-v0.1-300k/mscale.v0.1.300k.v1.0.n10.tem1.0.p1.0.prefer_pair.by_sc_p0.5.json \
--top_p 0.5
#Filtered 63546 samples.
#Total number of items: 298275
#Acc: 0.9530181613690681
#Pass at k: 0.7869549911993965
#No positive solutions: 0 / 298275
#No negative solutions: 156892 / 298275
#Num pairs: 1233371
python scripts/math_scale/construct_prefer_pair_sc.py \
--input_file "${OUTPUT_PREFIX_PATH}/experiments/mathstral.mathscale4o.process-dpo.iter0.V100.tp8dp48.v2.2.fix.s42/checkpoint-600/mathscale4o/mscale-v0.1-300k/mscale.v0.1.300k.v1.0.n10.tem1.0.p1.0.*-of-512.s42.json" \
--output_file ${OUTPUT_PREFIX_PATH}/experiments/mathstral.mathscale4o.process-dpo.iter0.V100.tp8dp48.v2.2.fix.s42/checkpoint-600/mathscale4o/mscale-v0.1-300k/mscale.v0.1.300k.v1.0.n10.tem1.0.p1.0.prefer_pair.by_sc_p0.0.json \
--top_p 0.0
#Filtered 0 samples.
#Total number of items: 298275
#Acc: 0.8771335177269298
#Pass at k: 1.0
#No positive solutions: 0 / 298275
#No negative solutions: 157701 / 298275
#Num pairs: 2538931
# Iter - 2
python scripts/math/deepseek_math_sample_steps.py \
--input_file "../msranlpintern/reward_modeling/experiments/mathstral.mscale-v0.1-300k.process-dpo-sc.p0.5.iter1.V100.tp4dp32.v3.1.s42/checkpoint-500/mathscale4o/mscale-v0.1-300k/mscale.v0.1.300k.v1.0.n10.tem1.0.p1.0.*-of-512.s42.json" \
--output_file ../msranlpintern/reward_modeling/experiments/mathstral.mscale-v0.1-300k.process-dpo-sc.p0.5.iter1.V100.tp4dp32.v3.1.s42/checkpoint-500/mathscale4o/mscale-v0.1-300k/mscale.v0.1.300k.v1.0.n10.tem1.0.p1.0.upper0.7.r0.3.sample10.filter_same.json \
--upper_step_ratio 0.7 --sample_ratio 0.3 --filter_all_same --sample_over_p 10
python scripts/math_scale/construct_prefer_pair_sc.py \
--input_file "${OUTPUT_PREFIX_PATH}/experiments/mathstral.mscale-v0.1-300k.process-dpo-sc.p0.5.iter1.V100.tp4dp32.v3.1.s42/checkpoint-500/mathscale4o/mscale-v0.1-300k/mscale.v0.1.300k.v1.0.n10.tem1.0.p1.0.*-of-512.s42.json" \
--output_file ${OUTPUT_PREFIX_PATH}/experiments/mathstral.mscale-v0.1-300k.process-dpo-sc.p0.5.iter1.V100.tp4dp32.v3.1.s42/checkpoint-500/mathscale4o/mscale-v0.1-300k/mscale.v0.1.300k.v1.0.n10.tem1.0.p1.0.prefer_pair.by_sc_p0.5.json \
--top_p 0.5
#Filtered 43373 samples.
#Total number of items: 298275
#Acc: 0.9614400828553719
#Pass at k: 0.8545872097896237
#No positive solutions: 0 / 298275
#No negative solutions: 192622 / 298275
#Num pairs: 986657
python scripts/math_scale/construct_prefer_pair_sc.py \
--input_file "${OUTPUT_PREFIX_PATH}/experiments/mathstral.mscale-v0.1-300k.process-dpo-sc.p0.5.iter1.V100.tp4dp32.v3.1.s42/checkpoint-500/mathscale4o/mscale-v0.1-300k/mscale.v0.1.300k.v1.0.n10.tem1.0.p1.0.*-of-512.s42.json" \
--output_file ${OUTPUT_PREFIX_PATH}/experiments/mathstral.mscale-v0.1-300k.process-dpo-sc.p0.5.iter1.V100.tp4dp32.v3.1.s42/checkpoint-500/mathscale4o/mscale-v0.1-300k/mscale.v0.1.300k.v1.0.n10.tem1.0.p1.0.prefer_pair.by_sc_p0.json \
--top_p 0.0
#Filtered 0 samples.
#Total number of items: 298275
#Acc: 0.9059827340541446
#Pass at k: 1.0
#No positive solutions: 0 / 298275
#No negative solutions: 193482 / 298275
#Num pairs: 1897089
# Iter - 3
python scripts/math/deepseek_math_sample_steps.py \
--input_file "../msranlpintern/reward_modeling/experiments/mathstral.mscale-v0.1-300k.process-dpo-sc.p0.0.iter2.H100.dp16.v1.3.s42/checkpoint-500/mathscale4o/mscale-v0.1-300k/mscale.v0.1.300k.v1.0.n10.tem1.0.p1.0.*-of-512.s42.json" \
--output_file ../msranlpintern/reward_modeling/experiments/mathstral.mscale-v0.1-300k.process-dpo-sc.p0.0.iter2.H100.dp16.v1.3.s42/checkpoint-500/mathscale4o/mscale-v0.1-300k/mscale.v0.1.300k.v1.0.n10.tem1.0.p1.0.upper0.7.r0.3.sample10.filter_same.json \
--upper_step_ratio 0.7 --sample_ratio 0.3 --filter_all_same --sample_over_p 10
python scripts/math_scale/construct_prefer_pair_sc.py \
--input_file "${OUTPUT_PREFIX_PATH}/experiments/mathstral.mscale-v0.1-300k.process-dpo-sc.p0.0.iter2.H100.dp16.v1.3.s42/checkpoint-500/mathscale4o/mscale-v0.1-300k/mscale.v0.1.300k.v1.0.n10.tem1.0.p1.0.*-of-512.s42.json" \
--output_file ${OUTPUT_PREFIX_PATH}/experiments/mathstral.mscale-v0.1-300k.process-dpo-sc.p0.0.iter2.H100.dp16.v1.3.s42/checkpoint-500/mathscale4o/mscale-v0.1-300k/mscale.v0.1.300k.v1.0.n10.tem1.0.p1.0.prefer_pair.by_sc_p0.0.json \
--top_p 0.0
PFPO/scripts/math_scale/mscale/rerank.sh
+15
View File
@@ -0,0 +1,15 @@
export OUTPUT_PREFIX_PATH=/mnt/fangkai_blob/reward_modeling/
target_dir=$1
#python scripts/math_scale/rerank_w_prm_math.py \
# --response_file "$OUTPUT_PREFIX_PATH/experiments/$target_dir/math/math.test.v1.1.0shot.n1.tem1.0.p1.0.0-of-1.s*[0-9].json" \
# --reward_file "$OUTPUT_PREFIX_PATH/experiments/mathstral.mscale-v0.1-300k.process-rm-sc.p0.0.iter2.V100.dp256.v1.0.s42/$target_dir/-*/test-checkpoint-500/eval_predictions.json" \
# --reduction min --num_workers 128 --re_index
python scripts/math_scale/rerank_w_prm_math.py \
--response_file "$OUTPUT_PREFIX_PATH/experiments/$target_dir/math/math.test.v1.1.0shot.n1.tem1.0.p1.0.0-of-1.s*[0-9].json" \
--reward_file "$OUTPUT_PREFIX_PATH/experiments/mathstral.mscale-v0.1-300k.process-rm-sc.p0.0.iter3.V100.dp256.v1.0.s42/$target_dir/-*/test-checkpoint-400/eval_predictions.json" \
--reduction min --num_workers 128 --re_index