irsotarriva/deepseek-v4.1-flash-top20-commitpackft

DeepSeek-V4.1-Flash top-20 log-probabilities on CommitPackFT prefixes

Teacher annotations for logit distillation on code. DeepSeek-V4.1-Flash continued 19,621 short windows of source files from bigcode/commitpackft. Each record has the continuation the teacher sampled and, at every generated position, the 20 most likely tokens with their log-probabilities.

The source code is not in this dataset. A record points to its prefix (source row, character offset, length, sha256) and rebuild.py fetches the text from CommitPackFT.

Two sister datasets were made the same way from other sources: prose and agent trajectories.

Numbers

	train	eval	all
records	18,621	1,000	19,621
generated tokens (positions with a top 20)	4,061,403	218,254	4,279,657

language	records
Python	8,298
JavaScript	5,026
TypeScript	2,407
C++	1,887
Go	1,071
Rust	932

• The top 20 hold 98.3% of the probability mass on average.
• 18,296 continuations ended because the teacher ended the text; 1,325 reached the 1,024-token limit.
• Mean length 218 tokens. 490 MB of Parquet.

How it was made

• Prefixes. From the post-commit file contents (new_contents) of six languages, in file order. A file shorter than 1,000 characters was skipped. From each other file, one window of 800 characters starting at a line boundary, at a random place.
• Teacher. DeepSeek API, model deepseek-flash (DeepSeek-V4.1-Flash), raw-completion endpoint, no chat template and no thinking, temperature 1, max_tokens 1024, logprobs 20. Run on 8 and 9 October 2026.
• Left out. Prefixes where the teacher ended the text at once (nothing to keep) and prefixes the provider's content filter refused.
• Splits. A file is in eval when a hash of its id falls in 1 of 20, otherwise in train.
• Token ids. The API returns token strings. They were mapped to ids of the DeepSeek-V4.1-Flash tokenizer. Every sampled token was mapped. 370 of 85.6M top-20 entries were not, and have id -1.

Fields

field	type	meaning
id	string	commitpackft/<lang>/<commit>/<new_file>
source_dataset	string	bigcode/commitpackft
source_revision	string	commit hash of the source dataset the prefix was read from
source_config	string	the language directory, data/<source_config>/data.jsonl
source_row	int64	line number in that file, from 0
prefix_start, prefix_length	int32	the window, in characters of new_contents
prefix_sha256	string	sha256 of the window, UTF-8
lang, commit, new_file, repos	string	copied from the CommitPackFT row
repo_license	string	the license field of the CommitPackFT row
finish_reason	string	stop (the teacher ended the text) or length
n_tokens	int32	generated tokens
tokens	list of string	the sampled tokens, as the API wrote them
token_ids	list of int32	their tokenizer ids
token_logprobs	list of float32	log-probability of each sampled token
top_ids	list of list of int32	the 20 most likely ids at each position, most likely first; -1 if unmapped
top_logprobs	list of list of float32	their log-probabilities

Log-probabilities are natural logarithms, rounded to 4 decimals. The sampled token is not always inside the top 20.

Getting the prefix text

pip install pyarrow
python rebuild.py "data/*.parquet" --out prefixes.jsonl

rebuild.py is in this repository. It reads the six CommitPackFT files at source_revision (about 310 MB), cuts the windows and checks each against prefix_sha256. The output has one id and text per record.

The training sequence of a record is: the start token (id 0), the tokenizer's ids for the prefix text, token_ids, and the end token (id 1) when finish_reason is stop. Only the token_ids positions have a teacher distribution.

Using it

• The top 20 are not the whole distribution. Put the remaining mass (2% on average here, more at uncertain positions) in a residual bucket, or renormalise.
• The positions are the teacher's own samples at temperature 1, not human text. The teacher is more confident on its own continuations than it would be on fixed text.
• The teacher often ends a file soon after the prefix (median below 100 tokens), so many records are short.

Licence and attribution

• The annotations (continuations, log-probabilities) are released under CC-BY-4.0. They are output of DeepSeek-V4.1-Flash obtained through the DeepSeek API. The DeepSeek Open Platform Terms assign the rights in outputs to the user and list training other models and distillation among the permitted uses. All text in tokens is AI-generated and may contain errors.
• The pointers and the copied columns come from CommitPackFT (Muennighoff et al., OctoPack: Instruction Tuning Code Large Language Models, 2023), released under the MIT licence.
• The prefixes are not distributed here. When you rebuild them you receive code under the licence of its repository, given per row in repo_license:

repo_license	records
mit	10,596
apache-2.0	3,965
bsd-3-clause	2,248
agpl-3.0	1,090
bsd-2-clause	667
lgpl-2.1	307
mpl-2.0	284
isc	180
unlicense	139
cc0-1.0	92
unknown	23
epl-1.0	16
artistic-2.0	14

8.8% of the records point to copyleft or unlicensed code (agpl-3.0, lgpl-2.1, mpl-2.0, epl-1.0, unknown). Filter on repo_license if that matters for your use.

• A continuation can repeat text of the source file when the teacher has memorised it.