Squish42's picture
Add 'pruned' configuration.
5a66577
name: Bluemoon-Split
description: "Splits"
build:
conversationBatch: 100
cleanDirectory: [ 'out/data/pruned', 'out/stats/pruned' ]
verifyRequirements: false
verifyArtifacts: false
sendPackedFiles: true
gzipRemotePackedFiles: true
gzipLocalPackedFiles: false
_input_steps:
- step: &bluemoon_cleaned
type: FastChatInput
sync: main
config:
path: "out/data/full/bluemoon.train.json"
preserveIds: true
_passes:
- pass: &cfg--pruned
required:
- path: "out/data/full/bluemoon.train.json"
steps:
# Pre-cleaned source
- *bluemoon_cleaned
# Filter
- type: StatsConversationPrune
description: "Prune conversations for deviation or maximum message length"
pack: [ 'out/stats/full/train/stats.json', 'out/stats/full/train/conversations.json.gz' ]
config:
packedStatsPath: 'out/stats/full/train'
wordsStdDevRatioMin: 0.5
#wordsStdDevRatioMax: 10.0
#wordsClampMin: 120
wordsClampMax: 2048
messagesCountMin: 4
# Train split
- type: FastChatOutput
sync: main
config:
path: "out/data/pruned/bluemoon.train.json"
indent: 2
- type: StatsOutput
description: "Output statistics"
sync: main
config:
path: 'out/stats/pruned/train/stats.json'
vocabPath: 'out/stats/pruned/train/vocabulary.json.gz'
gZipVocabulary: true
conversationsPath: 'out/stats/pruned/train/conversations.json.gz'
gZipConversations: true
# Test split
- type: RawSampleOutput
sync: main
config:
ratio: 0.01
forwardFilter: true
path: "out/data/pruned/bluemoon.test.raw"
- type: StatsOutput
description: "Output statistics"
sync: main
config:
path: 'out/stats/pruned/test/stats.json'
vocabPath: 'out/stats/pruned/test/vocabulary.json.gz'
gZipVocabulary: true
conversationsPath: 'out/stats/pruned/test/conversations.json.gz'
gZipConversations: true
artifacts:
- file: "out/data/pruned/bluemoon.train.json"
- file: "out/data/pruned/bluemoon.test.raw"
# Use the above pass as a template to output for new configurations and splits.
passes: [ *cfg--pruned ]