README.md

metadata

license: other
language:
  - en
  - tr
task_categories:
  - text-generation
tags:
  - tokenized
  - gemma-3
  - reasoning
  - sft
  - 4k

stage2_tr_polish_4k

Tokenized packed SFT dataset for google/gemma-3-1b-pt.

Format

• Split: train
• Columns: input_ids, length
• input_ids: packed 4096-token uint32 sequences
• Labels are not stored; use labels = input_ids.copy() in the causal LM collator.
• Assistant text before tokenization is normalized as:

<think>
reasoning
</think>
final answer

No extra special tokens were added to the tokenizer.

Stats

{
  "stage": "stage2_tr_polish_4k",
  "tokenizer": "google/gemma-3-1b-pt",
  "seq_len": 4096,
  "max_example_tokens": 16384,
  "target_tokens": 250000000,
  "written_tokens": 250003456,
  "resumed_from_tokens": 0,
  "num_sequences": 61036,
  "num_shards": 8,
  "shard_rows": 8192,
  "parquet_zstd_level": 1,
  "repo_id": "Efe2898/gemma3-1b-reasoning-stage2-tr-polish-4k-tokenized",
  "streaming": true,
  "cpu_cores": 4,
  "tokenize_batch_size": 2048,
  "source_batch_size": 2048,
  "weights": {
    "NuminaMath-CoT": 5,
    "OpenR1-Math-220k": 5,
    "NuminaMath-TIR": 2,
    "OpenThoughts-114k": 3,
    "turkish-math-186k": 40,
    "gsm8k_tr": 12,
    "OpenThoughts-TR-18k": 23,
    "turkish-general-reasoning-28k": 10
  },
  "splits": {
    "NuminaMath-CoT": "train",
    "OpenR1-Math-220k": "train",
    "NuminaMath-TIR": "train",
    "OpenThoughts-114k": "train",
    "turkish-math-186k": "train",
    "gsm8k_tr": "train",
    "OpenThoughts-TR-18k": "train",
    "turkish-general-reasoning-28k": "train"
  },
  "produced_tokens_by_source_approx": {
    "NuminaMath-CoT": 10848666,
    "OpenR1-Math-220k": 12033160,
    "NuminaMath-TIR": 3071805,
    "OpenThoughts-114k": 17651159,
    "turkish-math-186k": 97819415,
    "gsm8k_tr": 27835316,
    "OpenThoughts-TR-18k": 55358541,
    "turkish-general-reasoning-28k": 25385720
  },
  "examples_by_source_approx": {
    "turkish-math-186k": 174080,
    "OpenThoughts-TR-18k": 14336,
    "gsm8k_tr": 137216,
    "turkish-general-reasoning-28k": 12288,
    "NuminaMath-CoT": 21823,
    "OpenR1-Math-220k": 2048,
    "OpenThoughts-114k": 2048,
    "NuminaMath-TIR": 4096
  },
  "skipped_too_long_by_source": {},
  "note": "Very long examples are truncated inside tokenizer at max_example_tokens for speed.",
  "disabled_sources_during_this_run": {},
  "active_sources_at_end": [
    "NuminaMath-CoT",
    "OpenR1-Math-220k",
    "NuminaMath-TIR",
    "OpenThoughts-114k",
    "turkish-math-186k",
    "gsm8k_tr",
    "OpenThoughts-TR-18k",
    "turkish-general-reasoning-28k"
  ]
}

Source mix

Source	Weight	Repo	Split(s)
NuminaMath-CoT	5	AI-MO/NuminaMath-CoT	train
OpenR1-Math-220k	5	open-r1/OpenR1-Math-220k	train
NuminaMath-TIR	2	AI-MO/NuminaMath-TIR	train
OpenThoughts-114k	3	open-thoughts/OpenThoughts-114k	train
turkish-math-186k	40	ituperceptron/turkish-math-186k	train
gsm8k_tr	12	ytu-ce-cosmos/gsm8k_tr	train
OpenThoughts-TR-18k	23	selimc/OpenThoughts-TR-18k	train
turkish-general-reasoning-28k	10	ituperceptron/turkish-general-reasoning-28k	train

License note

This is a mixed-source tokenized derivative dataset. Check each upstream dataset license/provenance before public or commercial use.