Datasetpublic

kuluruvineeth/manas_dataset

Manas Dataset The complete English training bundle for Manas, a lightweight language model trained entirely from scratch. Every file is rebuildable from raw sources with python -m datapipe.build all. Files file stage pretrain_t2t.jsonl pretraining corpus, ~2.2B tokens pretrain_t2t_mini.jsonl pretraining corpus, quick-start tier sft_t2t.jsonl supervised fine-tuning conversations (tool-calling and reasoning mixed in) sft_t2t_mini.jsonl supervised… See the full description on the dataset page: https://huggingface.co/datasets/kuluruvineeth/manas_dataset.

sourceHugging Faceotherupdated 5d agoView on Hugging Face

0likes63downloads

Dataset Card

Manas Dataset

The complete English training bundle for Manas, a lightweight language model trained entirely from scratch. Every file is rebuildable from raw sources with python -m datapipe.build all.

Files

file	stage
`pretrain_t2t.jsonl`	pretraining corpus, ~2.2B tokens
`pretrain_t2t_mini.jsonl`	pretraining corpus, quick-start tier
`sft_t2t.jsonl`	supervised fine-tuning conversations (tool-calling and reasoning mixed in)
`sft_t2t_mini.jsonl`	supervised fine-tuning, quick-start tier
`dpo.jsonl`	preference pairs for DPO
`rlaif.jsonl`	prompt-only rows for PPO/GRPO
`agent_rl.jsonl`	synthesized tool-use RL tasks with verifiable ground truth
`agent_rl_math.jsonl`	GSM8K word problems with numeric ground truth
`lora_identity.jsonl`	Manas identity adapter data
`lora_medical.jsonl`	medical-domain adapter data
`lora_exam.jsonl`	exam-format alignment data

Sources

Every upstream source, its license as stated on its dataset card, and why it was chosen:

source	license	why
`HuggingFaceFW/fineweb-edu` (sample-10BT)	odc-by	quality-filtered educational web text; the strongest general English base for a small model
`HuggingFaceTB/smollm-corpus` (cosmopedia-v2)	odc-by	synthetic textbook-style articles; dense knowledge per token, ideal at sub-100M scale
`roneneldan/TinyStories`	cdla-sharing-1.0	simple narratives that give tiny models early grammatical coherence
`HuggingFaceTB/smoltalk` (all)	not stated on card (HuggingFaceTB release)	1M multi-turn conversations curated specifically for sub-1B models
`teknium/OpenHermes-2.5`	not stated on card (compilation of variously-licensed sets)	1M diverse instruction conversations; breadth of tasks and styles
`HuggingFaceH4/ultrachat_200k`	mit	200k long multi-turn dialogues; strengthens conversational depth
`NousResearch/hermes-function-calling-v1` (func_calling)	apache-2.0	tool-calling conversations already using <toolcall>/<toolresponse> tags
`NousResearch/hermes-function-calling-v1` (funccallingsingleturn)	apache-2.0	single-turn tool-calling; clean minimal examples of the call format
`NousResearch/hermes-function-calling-v1` (glaivefunccalling)	apache-2.0	curated glaive subset in hermes tagging; adds tool diversity
`glaiveai/glaive-function-calling-v2`	apache-2.0	113k function-calling dialogues; the volume backbone of tool-use training
`open-thoughts/OpenThoughts-114k`	apache-2.0	reasoning traces split into reasoning_content + answer; only short traces fit a small context
`HuggingFaceH4/ultrafeedback_binarized`	mit	canonical chosen/rejected preference pairs; schema-identical to our DPO contract
`lavita/ChatDoctor-HealthCareMagic-100k`	not stated on card (mirror of ChatDoctor data)	real patient-question medical QA for the domain-adapter LoRA demo
`openai/gsm8k` (main)	mit	grade-school word problems with verifiable numeric answers; ground truth for agentic RL
`cais/mmlu` (all)	mit	exam-format alignment; validation split only, never test
`cais/mmlu` (all)	mit	exam-format alignment; dev split only, never test
`allenai/ai2_arc` (ARC-Easy)	cc-by-sa-4.0	science multiple choice; train split only
`allenai/ai2_arc` (ARC-Challenge)	cc-by-sa-4.0	harder science multiple choice; train split only
`allenai/openbookqa` (main)	unknown (per dataset card)	open-book science questions; train split only

agent_rl.jsonl is synthesized by 6 deterministic generators over fixed mock-tool tables, so every ground-truth answer is correct by construction.

No source uses a benchmark test split.