datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
pjm-demand-weather
PJM Hourly Electricity Demand + Population-Weighted Weather (2015 to 2026)
A model ready hourly time series for forecasting load on the largest US
grid: PJM RTO demand (MWh) pre joined to population weighted weather,
with calendar, holiday, degree hour, irradiance, cloud cover, apparent
temperature, and snowfall features. 97,850 hours from 2015-07-01 to
2026-08-29, no missing hours on the UTC spine, every imputed,
interpolated, or preliminary value explicitly flagged, in Parquet… See the full description on the dataset page: https://huggingface.co/datasets/Arimancy/pjm-demand-weather.proxy-logs-ReRolls-Minos
non-refusal responses: 845,186
refusal responses: 38,285
https://gist.github.com/xzuyn/1d7f43db2750060a18a304eb84b396db
Use a training prompt formatter like this: https://github.com/xzuyn/axolotl/blob/latest-formatters/src/axolotl/prompt_strategies/customllama3-regex-last-only-prefill-reroll.py
proxy-logs-ReRollsDuplicate prompts combined into a single sample, with all responses in a list of dicts. I've also included some info like token count, and slop (though my slop list could use improvement).
Use a training prompt formatter like this: https://github.com/xzuyn/axolotl/blob/84aec029dfa9eb9670b8a51d432a279be6c85871/src/axolotl/prompt_strategies/customllama3-regex-last-only-prefill-reroll.py
Dataset creation script: https://gist.github.com/xzuyn/aa1f30b7394d2997766bef82edb67227
dolphin-deepseek-1k-think-1k-response-filtered-ShareGPTKodCode_KodCode-V1-SFT-R1-4k-think-1k-response-ShareGPTdolphin-flash-1k-think-1k-response-filtered-ShareGPTMedical-R1-Distill-Data-1k-think-512-response-filtered-ShareGPTnvidia-r1-code-1k-think-256-response-filtered-ShareGPToumi-ai_lmsys_chat_1m_clean_R1-1k-think-1k-response-ShareGPTSubtitles-rag-answers-r1
Subtitles-rag-answers-r1
You should mask everything except the last turn. The only part that matters to teach the model is the last turn, as you are teaching it to always output thinking, no matter what the user feeds it.
It's setup to be trained like R1:
Subtitles-rag-questions-r1
Subtitles-rag-questions-r1
You should mask everything except the last turn. The only part that matters to teach the model is the last turn, as you are teaching it to always output thinking, no matter what the user feeds it.
It's setup to be trained like R1:
HuggingFaceFW_finewiki-en-shuffledargilla_ultrafeedback-multi-binarized-quality-preferences-cleaned-PreferenceShareGPTFizzarolli_FallingThroughTheSkies-592k-Filtered-Filtered-subsetmedical-o1-reasoning-SFT-ENG-ShareGPTThinkyThinky-PJ-Only-Gemini-ShareGPTvarious-rp-sets-qwqargilla_distilabel-math-preference-dpo-PreferenceShareGPTHuggingFaceFW_fineweb-edu-subsetLLaMa-3 Token Count: 100,756,997
OpenThoughts-114k-Code_decontaminated-4k-think-2k-response-filtered-ShareGPTfoundRP-qwq-all-kcpp
foundRP-qwq-all-kcpp
You should mask everything except the last turn. The only part that matters to teach the model is the last turn, as you are teaching it to always output thinking, no matter what the user feeds it.
It's setup to be trained like R1:
goodwiki-2024-12-04argilla_ultrafeedback-binarized-preferences-cleaned-PreferenceShareGPTMagpie-Align_Magpie-Pro-DPO-200K-PreferenceShareGPTLMSYS-Chat-1M-ReRolls-Skywork-Reward-V2ThinkyThinky-PJ-Only-Gemini-2k-think-4k-response-filtered-ShareGPTNyxKrage_chub-logs-sharegpt-longest-CustomShareGPTPJMixers-Dev__LLaMa-3.1-Instruct-Interleaved-Zeroed-13Baesir-rpg-fantasy-novel-flatguard-split-qwqLMSYS-Chat-1M-ReRolls-Skywork-Reward-V2-Minos
non-refusal responses: 188838
refusal responses: 58766