enseven/kodcode-lfm2.5

KodCode LFM2.5

A preprocessed version of KodCode-V1-SFT-R1 formatted for fine-tuning LFM2.5-1.2B-Thinking with a 70% code-only / 30% CoT (chain-of-thought) mix.

Dataset Summary

This dataset is derived from KodCode-V1-SFT-R1 (CC BY-NC 4.0). For each of the 268,211 training rows, the question-answer pair is formatted into a flat chat template string using <|im_start|> / <|im_end|> markers and a <|startoftext|> prefix — the native format for LFM2.5-1.2B-Thinking.

Each row is assigned to one of two formats:

• code-only (70%): question → r1_solution (code response, no thinking trace)
• CoT (30%): question → conversations[-1]["value"] (full response with <think> reasoning)

The CoT selection is limited to texts ≤ 15,000 characters (guaranteed to fit within 4096 tokens at the observed minimum char-to-token ratio of 2.70).

Changes from the Original

1. Format conversion: Original conversations field (list of from/value dicts) is flattened into a single text string with <|im_start|> / <|im_end|> chat template.
2. Format selection: Each row is assigned either code-only or cot format at a 70/30 ratio (controlled random per shard).
3. Char pre-filtering: CoT texts exceeding 15,000 characters fall back to code-only format (instead of being dropped), since only ~50% of CoT texts fit within 4096 tokens.
4. Reduced columns: Only 4 columns are kept: text, format, subset, question_id.

Data Fields

Field	Type	Description
text	string	Flattened chat text: <|startoftext|><|im_start|>user\n{question}<|im_end|>\n<|im_start|>assistant\n{answer}<|im_end|>\n
format	string	"code-only" or "cot"
subset	string	Original KodCode subset (e.g. "Leetcode", "Codeforces", "Taco", etc.)
question_id	string	Original question identifier from KodCode

Column mapping to original dataset

This dataset	KodCode-V1-SFT-R1
text (code-only)	question + r1_solution formatted with chat template
text (CoT)	question + conversations[-1]["value"] formatted with chat template
subset	subset
question_id	question_id
(omitted)	solution, test, test_info, version, style, metadata, r1_pass_sequence, r1_correctness, gpt_pass_sequence, gpt_difficulty, gpt_pass_percentage, conversations

Data Splits

Split	Size
train	268,211 rows

Measured outcome when used for SFT

This dataset mix was used to LoRA fine-tune LFM2.5-1.2B-Thinking (adapter). On a sealed 128-task HumanEval+ evaluation, the fine-tuned model scored below the base model it was trained from (paired plus-pass 49 vs 56; -7.1 pp). The mechanism is instructive: the 70% code-only r1_solution targets taught a reasoning-capable base to skip its reasoning traces (output length collapsed ~11x), and correctness fell with them, while output formatting improved.

Recommendations for downstream use:

• Preserve CoT for reasoning-capable bases. The 30% CoT slice did not compensate for the 70% code-only targets. Train on <think>-style solutions for code, or distill the base model's own verified traces.
• Execution-verify training targets. KodCode-V1-SFT-R1's r1_solution answers were treated as ground truth here; filter to solutions that pass their unit tests, and exclude KodCode's shipped incorrect subset.
• Full evaluation evidence: adapter model card and GitHub — lfm2.5-finetune-code (reports/e3/, reports/e4/).

Usage

Load with HuggingFace Datasets

from datasets import load_dataset

ds = load_dataset("kodcode_dataset", split="train")
# or from parquet directly:
ds = Dataset.from_parquet("kodcode_dataset/data/kodcode-lfm2.5.parquet")

Load for SFT training (TRL)

from trl import SFTTrainer

trainer = SFTTrainer(
    ...,
    train_dataset=ds,
    dataset_text_field="text",
    max_seq_length=4096,
    packing=True,
)

Subset distribution

print(ds.to_pandas()["subset"].value_counts())

Format distribution

print(ds.to_pandas()["format"].value_counts())
# code-only    187824
# cot           80387

Statistics

Metric	Value
Total rows	268,211
Code-only rows	187,824 (70.0%)
CoT rows	80,387 (30.0%)
Total characters	~1.09B
Text length (mean)	4,046 chars
Text length (median)	2,130 chars
Text length (max)	15,000 chars
Output size (parquet)	442 MB

Citation

If you use this dataset, please cite the original KodCode work:

@article{xu2025kodcode,
      title={KodCode: A Diverse, Challenging, and Verifiable Synthetic Dataset for Coding},
      author={Zhangchen Xu and Yang Liu and Yueqin Yin and Mingyuan Zhou and Radha Poovendran},
      year={2025},
      eprint={2503.02951},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2503.02951},
}

License

This dataset is derived from KodCode-V1-SFT-R1 and is distributed under the same CC BY-NC 4.0 license.

• Attribution: You must give appropriate credit to the original KodCode authors.
• NonCommercial: You may not use the material for commercial purposes.
• No additional restrictions: You may not apply legal terms that restrict others from doing anything the license permits.

See the full license text for details.