algorithmic-sft
algorithmic-sft-training-data-v1
algorithmic-sft-training-data-v1
Algorithmic SFT training data: deterministic step-by-step traces for 5 domains (countdown, formal_logic, long_arithmetic, conlang_morphology, cellular_automata) across multiple algorithm variants. Programmatically generated — no LLM involved.
Dataset Info
Rows: 63000
Columns: 8
Columns
Column
Type
Description
question
Value('string')
The problem statement presented to the model
answer
Value('string')
The correct… See the full description on the dataset page: https://huggingface.co/datasets/reasoning-degeneration-dev/algorithmic-sft-training-data-v1.algorithmic-sft-training-configs-v1
algorithmic-sft-training-configs-v1
LlamaFactory training configs. All cutoff_len=32768. Countdown configs use new equation-answer format.
Dataset Info
Rows: 17
Columns: 7
Columns
Column
Type
Description
config_name
Value('string')
YAML filename
domain
Value('string')
No description provided
is_distillation
Value('bool')
No description provided
yaml_content
Value('string')
Full YAML config
model_name
Value('string')
No description provided… See the full description on the dataset page: https://huggingface.co/datasets/reasoning-degeneration-dev/algorithmic-sft-training-configs-v1.algorithmic-sft-full-eval-v4
algorithmic-sft-full-eval-v4
Aggregate eval results: 10 models x 4 domains x 3 splits with bootstrap 95% CIs
Dataset Info
Rows: 42
Columns: 8
Columns
Column
Type
Description
model
Value('string')
HuggingFace model ID (LoRA adapter name)
domain
Value('string')
Task domain: formal_logic, conlang_morphology, cellular_automata, long_arithmetic
type
Value('string')
Training type: algo (algorithmic SFT) or distill (QwQ distillation)
split… See the full description on the dataset page: https://huggingface.co/datasets/raca-workspace-v1/algorithmic-sft-full-eval-v4.algorithmic-sft-sharegpt-training-v1
algorithmic-sft-sharegpt-training-v1
Exact LlamaFactory training data. Countdown uses equation-answer format with Step 1:.... Other domains use Answer: X. All wrapped in tags.
Dataset Info
Rows: 82903
Columns: 4
Columns
Column
Type
Description
conversations
List({'from': Value('string'), 'value': Value('string')})
ShareGPT — literal LlamaFactory input
source_file
Value('string')
JSON filename → dataset_info.json
model_type
Value('string')… See the full description on the dataset page: https://huggingface.co/datasets/reasoning-degeneration-dev/algorithmic-sft-sharegpt-training-v1.algorithmic-sft-eval-sets-v1
algorithmic-sft-eval-sets-v1
Evaluation sets for algorithmic SFT experiment: test (1000/domain), val (200/domain), harder variant (500/domain), structural OOD (500/domain). 5 domains total.
Dataset Info
Rows: 11000
Columns: 9
Columns
Column
Type
Description
question
Value('string')
The problem statement
answer
Value('string')
The correct answer
sft_trace
Value('null')
Reference algorithmic trace (golden solution)
difficulty
Value('int64')… See the full description on the dataset page: https://huggingface.co/datasets/reasoning-degeneration-dev/algorithmic-sft-eval-sets-v1.algorithmic-sft-distillation-training-data-v1
algorithmic-sft-distillation-training-data-v1
QwQ-32B distillation training data for 5 algorithmic domains. Correct responses filtered by collect_distill_results.py with truncation rejection. KNOWN ISSUE: countdown domain has 74.2% QwQ repetition loops (model repeats \boxed{} answer until hitting token limit). Countdown is being regenerated with v3 pipeline (32k tokens). Other 4 domains are clean (>99% quality).
Dataset Info
Rows: 24133
Columns: 6
Columns… See the full description on the dataset page: https://huggingface.co/datasets/reasoning-degeneration-dev/algorithmic-sft-distillation-training-data-v1.