udgamlabs/laghu-sft

UDGAM LABS · Dataset · 40,853 of the 42,320 conversations Laghu was fine-tuned on

Laghu SFT data

These are the supervised fine-tuning (SFT) conversations that turned Laghu-Base-32K into Laghu, Udgam Labs' 152M-parameter English chat model. Laghu was trained on 41,467 training and 853 validation conversations. This release has 40,032 + 821 of them, exactly as trained: when re-rendered with Laghu's chat template, every row with unchanged text gives the same token ids that were trained on. The 1,467 rows built on MATH competition problems are left out (ids and a rebuild script only), and the text of 351 FinePDFs-Edu documents is replaced by a placeholder that a script fills back in. Every row says where its prompt came from, which model wrote each part, and its licence.

Model: Laghu · code on GitHub · udgamlabs.com/models/laghu

Quick start

import json
from datasets import load_dataset
from transformers import AutoTokenizer

ds = load_dataset("udgamlabs/laghu-sft")
tok = AutoTokenizer.from_pretrained("udgamlabs/Laghu")

def to_messages(cells):
    """Rows store each message as a fixed struct; drop empty fields and parse tool calls back from JSON."""
    out = []
    for c in cells:
        m = {k: v for k, v in c.items() if v is not None}
        if "tool_calls" in m:
            m["tool_calls"] = json.loads(m["tool_calls"])
        out.append(m)
    return out

row = ds["train"][0]
text = tok.apply_chat_template(to_messages(row["messages"]),
                               tools=json.loads(row["tools"]) if row["tools"] else None,
                               tokenize=False, enable_thinking=row["enable_thinking"])
print(row["label"])
print(text)

This is how the training text was made. Loss was on the model turns only: the tokens after <|turn>model\n up to and including <turn|>, with tool results masked except the <|tool_response> hand-over token (rule in udgam/sft/render_sft.py on GitHub). epochs says how often a row appeared in one training pass (2 for identity rows, otherwise 1); Laghu used 2 passes.

To get the FinePDFs-Edu document text and the MATH prompts back:

python rebuild_text.py --release . --out laghu-sft-rebuilt

What's inside

Rows in this release, by slice (a slice is one kind of training conversation):

Slice	Train	Validation	Thinking on	What the rows teach
chat	8,232	157	0	everyday requests from real users; 2,316 train rows have 2-3 user turns
chat_thinking	1,471	29	1,471	the same kind of requests, with a short visible thinking step
gen_mt	10,388	217	0	multi-turn follow-ups ("shorter", "and for dogs?", "why?")
instruction_following	3,743	80	0	prompts with checkable constraints (length, format, keywords)
system_prompts	1,469	31	0	following an operator's system prompt or persona
maths_logic	7,333	157	2,444	word problems, logic puzzles and program tracing, answer checked; thinking-on rows have a thinking-off twin
python_tool	2,946	54	1,418	writing a Python call, reading its real output, answering
json_function_calls	1,473	27	601	calling tools from JSON schemas, including parallel, multi-step and no-call cases
long_documents	1,404	30	0	summaries and questions over 8K-14K-token documents
safety	1,173	27	0	declining harmful requests briefly, and answering borderline-but-fine ones normally
identity	400	12	98	who the model is (templated)
Total	40,032	821	6,032	14,290 train rows (35.7%) have two or more user turns

The trained mix also had 1,435 + 32 maths_logic rows on MATH-train problems; they are listed in math_train_ids.jsonl and not included here.

Sources and licences

Each row has a row_licence. "A AND B" means both licences apply to the row: our parts (generated answers, templates, our prompts) are CC-BY-4.0, and the upstream text keeps its own licence.

Prompt source	Train / val rows	Prompt licence	row_licence	Who wrote the text
WildChat-4.8M (AI2)	7,001 / 144	ODC-By-1.0	CC-BY-4.0 AND ODC-By-1.0	prompts: real users; answers: DeepSeek-V4.1-Flash
Arena human-preference 100k (LMArena)	5,166 / 95	CC-BY-4.0 (user prompts)	CC-BY-4.0	real users; DeepSeek-V4.1-Flash
Arena human-preference 140k (LMArena)	868 / 17	CC-BY-4.0 (user prompts)	CC-BY-4.0	real users; DeepSeek-V4.1-Flash
oasst2 (OpenAssistant)	876 / 16	Apache-2.0	CC-BY-4.0 AND Apache-2.0	volunteers; DeepSeek-V4.1-Flash
Wikipedia 20231101.en	585 / 12	CC-BY-SA-4.0	CC-BY-SA-4.0	Wikipedia contributors (title and URL per row); DeepSeek-V4.1-Flash
FinePDFs-Edu (Hugging Face)	1,028 / 23	ODC-By-1.0	CC-BY-4.0 AND ODC-By-1.0	PDF authors (text omitted); summaries and answers: DeepSeek-V4.1-Flash
GSM8K train (OpenAI)	1,908 / 41	MIT	CC-BY-4.0 AND MIT	dataset authors; DeepSeek-V4.1-Flash
GSM8K train with one added sentence	434 / 7	MIT AND CC-BY-4.0	CC-BY-4.0 AND MIT	the added distracting sentence: OpenAI gpt-6.1-sol (153 / 3), Cursor Composer 2.5 (130 / 4), xAI Grok 4.7 (151 / 0); answers: DeepSeek-V4.1-Flash
Practice problems from the Udgam Labs data factory	6,712 / 138	CC-BY-4.0	CC-BY-4.0	problems and programs: Cursor Composer 2.5 (5,388 / 99), OpenAI gpt-6.1-sol (904 / 30), xAI Grok 4.7 (420 / 9); answers: DeepSeek-V4.1-Flash
Logic puzzles generated by our code	1,023 / 22	CC-BY-4.0	CC-BY-4.0	puzzles: Udgam Labs Python generators (no model text); answers: DeepSeek-V4.1-Flash
Prompts generated by DeepSeek-V4.1-Flash (tool schemas and requests, safety requests, operator system prompts, Python-tool requests)	3,643 / 77	CC-BY-4.0	CC-BY-4.0	DeepSeek-V4.1-Flash
First exchanges from Nemotron-Post-Training-Dataset-v2 chat (NVIDIA)	5,675 / 123	CC-BY-4.0	CC-BY-4.0	first exchange: NVIDIA's data (replies by Qwen3-235B-A22B / Qwen3-30B-A3B); follow-ups and replies: gpt-oss-120b
First exchanges from SYNTH (Pleias)	2,809 / 51	CC-BY-4.0	CC-BY-4.0	first exchange: SYNTH (replies by Pleias' Qwen3-8B fine-tunes); follow-ups and replies: gpt-oss-120b
Conversations written whole by gpt-oss-120b	1,904 / 43	CC-BY-4.0	CC-BY-4.0	gpt-oss-120b
Identity rows	400 / 12	CC-BY-4.0	CC-BY-4.0	templates by Udgam Labs (not model output)
MATH train (Hendrycks et al.)	0 (1,435 / 32 left out)	not re-published	n/a	see "What was left out or changed"

Answer generators. DeepSeek-V4.1-Flash (through the DeepSeek API, model deepseek-flash, 2026-10-17) wrote the answers in 29,244 train / 592 validation rows. gpt-oss-120b (OpenAI, Apache-2.0), which we ran ourselves with vLLM on 2026-10-17, wrote the follow-up turns of 10,388 / 217 gen_mt rows. The 412 identity rows are templates. In python_tool rows, the tool output is the real output of running the code.

Columns

Column	Meaning
id	row id, the same as in the trained data (-think / -nothink mark the two thinking modes of one maths prompt)
slice, category	the kind of conversation and its sub-type
enable_thinking	rendered with thinking on (the reasoning_content of model turns becomes the thought channel)
messages	list of {role, content, reasoning_content, tool_calls, tool_call_id, name}; unused fields are null; tool_calls is a JSON string
tools	the tool definitions shown to the model (JSON string), or null
n_tokens, n_loss_tokens, epochs	token counts of the trained row (FinePDFs rows: with the document text), and passes per epoch
source_dataset, source_id, source_revision	where the prompt came from: dataset, upstream id and pinned commit
source_url, source_title	per-item page or document (Wikipedia and FinePDFs-Edu rows)
followup_source	chat rows: whether the second user turn is the same real user's (real user) or written by DeepSeek-V4.1-Flash (generated)
prompt_licence	licence of the user-side text
problem_author	data-factory rows: the model (or our code) that wrote the problem
answer_generator	the model that wrote the answers; seed_generator: the model behind a gen_mt row's first reply
row_licence	licence of the row (SPDX-style; "A AND B" means both apply)
ai_generated, label	whether the row contains AI-generated text, and a plain-language label naming each model
created_with_grok	xAI Grok 4.7 wrote text in this row (its label then says "Created with Grok")
verification	how the row was checked when it was made (for example "final answer == gold")
text_omitted, text_omitted_reason	the document text was replaced by a placeholder (FinePDFs-Edu rows)
pii_redactions	placeholders inserted in this row, if any (see below)

Ids in source_id: WildChat rows use conversation_hash:turn; oasst2 rows use message ids; GSM8K rows use gsm8k-train-<index>; data-factory rows use the factory's item id (the factory dataset itself is not public); gen_mt rows use the position of the seed conversation in our filtered seed pool, because the upstream row ids were not kept.

AI-generated content

All rows except the 412 identity rows contain AI-generated text, and each row's label names the models, for example "AI-generated by DeepSeek-V4.1-Flash (answer); word problem written by Cursor Composer 2.5". 580 rows (571 train, 9 validation) contain text written by xAI Grok 4.7 and are marked "Created with Grok". Answers can be wrong, even where a check passed: the checks cover final answers, code output, constraints and JSON validity, not every sentence.

The data-factory problems were written with OpenAI gpt-6.1-sol, Cursor Composer 2.5 and xAI Grok 4.7. Those services' terms bound Udgam Labs as the account holder; they are not passed on to you through this dataset. If you generate more data with those services, their terms apply to you directly. DeepSeek, OpenAI, Anysphere (Cursor), xAI, NVIDIA, Pleias and the dataset authors named here are credited as sources; none of them endorses this dataset.

What was left out or changed

• MATH train problems (1,435 train / 32 validation rows): left out. The original MATH dataset (hendrycks/competition_math) is disabled on Hugging Face after a DMCA takedown notice, so we do not re-publish its problems. math_train_ids.jsonl lists every row with its upstream pointer (EleutherAI/hendrycks_math, subject, index, pinned revision), our prompt wrapper and a sha256 of the exact prompt. rebuild_text.py rebuilds the prompts and checks the hashes. Our answers to these problems are not included either, so these rows rebuild as prompts only.
• FinePDFs-Edu documents (1,028 / 23 rows, 351 documents): text replaced by a placeholder. Each PDF stays its author's copyright. The rows keep our summaries and question-answer turns (answers sometimes quote a short passage), the title, URL, document id and revision. finepdfs_docs.jsonl gives each document's file and row in HuggingFaceFW/finepdfs-edu at the pinned revision and a sha256 of the exact text; rebuild_text.py downloads it (one row group of about 12 MB per document), cleans it the same way and checks the hash.
• Personal data and credential-like strings: replaced by placeholders. A rule-based scan of every string (emails, phone numbers, IP addresses, Luhn-checked card numbers, US SSNs, Indian PAN and Verhoeff-checked Aadhaar numbers, and API-key patterns) found 71 real-looking hits in the conversation text of 38 rows, now [EMAIL] (19), [PHONE] (22), [IP_ADDRESS] (25), [AADHAAR] (2) and [SECRET] (3). The URLs of 2 FinePDFs-Edu documents (6 rows) had an IP address or a token-like string, also replaced. Reserved example domains, private IP ranges, fictional 555 numbers and published test card numbers were left as they are. Counts by type and source are in SCAN_REPORT.md; no matched text is shown anywhere. The 38 rows differ from the trained text only at those placeholders.

How it was made

• Prompts. Real-user prompts from WildChat-4.8M, Arena and oasst2 (pinned revisions), filtered for English, personal data, unsafe topics and requests that need the web, files or images. Documents from Wikipedia and FinePDFs-Edu, screened for extraction errors and private persons. Problems from GSM8K train and our data factory, whose items were checked by programs (sandboxed code runs, gold answers, brute-force solvers, or an independent solve by a different model). Tool schemas, safety requests and operator system prompts generated by DeepSeek-V4.1-Flash.
• Answers. DeepSeek-V4.1-Flash, asked for short, structured answers and short thinking (300 words or fewer). Automatic checks per slice: answers equal the gold or the program output, the thinking and both replies state the same result, constraints are met (checked by code), JSON parses and matches the schema, code runs in a sandbox, and document answers are backed by a verbatim quote. Every kept row also passed filters for English output, personal data in the answer, mentions of the teacher model, and overlap with evaluation data: user turns sharing an 8-gram with MMLU, ARC, HellaSwag, GSM8K test, PIQA, WinoGrande, OpenBookQA, CommonsenseQA, IFEval, MATH-500, BFCL or our held-out chat prompts were dropped.
• Multi-turn rows. gpt-oss-120b continued first exchanges from Nemotron-Post-Training-Dataset-v2 and SYNTH with short follow-ups that depend on the earlier turns, and wrote whole conversations that start with a vague request. Rows naming our earlier model Gargi were dropped.
• Rendering. Laghu's chat template (Google's Gemma 4 template, Apache-2.0), up to 16,384 tokens per row; 2% of rows went to validation by a hash of the row id.
• This release. build_sft_release.py re-creates the trained rows from the same inputs, joins the data factory's records to name each problem's author, and applies the omissions and redactions above. verify re-renders every shipped row and compares it with the trained data: all 39,764 rows with unchanged text give identical token ids and loss masks, and with the FinePDFs-Edu text re-inserted by rebuild_text.py, all 40,815 rows without a redaction do (details in SCAN_REPORT.md).

Where it falls short

• Incomplete without the scripts. 1,467 trained rows (3.5%) are missing, and the MATH rows cannot be rebuilt in full because their answers are not included. The 1,051 FinePDFs-Edu rows need rebuild_text.py and a download of up to 351 row groups.
• The validation split leaks. It was split by row id, so related rows (the thinking on/off twins of one prompt, several questions on one document, multi-turn rows from one seed) sit on both sides: about 37% of eligible validation conversations had a near-copy in training. Do not use it as a clean held-out set.
• Two teachers, their errors included. Almost every answer comes from DeepSeek-V4.1-Flash or gpt-oss-120b. Open chat answers were screened by rules, not checked for correctness, and carry those models' mistakes and biases.
• Written for a 152M model. Answers are short (usually under about 250 words) and thinking is brief. That suits small models; it is not a general-purpose SFT set for larger ones.
• English only, and the safety slice is small (1,200 rows).
• The personal-data scan is rule-based. It looks for the patterns listed above, not for names or street addresses. Real-user prompts also went through their datasets' own de-identification (WildChat) and our filters, but some personal details may remain. Please report any you find.
• Seed ids. gen_mt rows point to positions in our filtered seed pool, not to upstream row ids, and their seed datasets were not pinned to a revision.
• Factory items are not public. Data-factory source_id values refer to our private factory dataset.
• 38 rows differ from the trained text at their redaction placeholders.

Removal requests

If a row contains text you wrote and you want it removed, for example a WildChat or Arena prompt, open a discussion on this dataset's Hugging Face page with the row's source_id (for WildChat, the conversation_hash:turn value). Do not post the personal details themselves. We will remove the matching rows in the next version. WildChat's own removal process for the upstream dataset is described on the WildChat-4.8M card.

Credits

We thank the authors and maintainers of the sources: AI2 (WildChat-4.8M; Zhao et al., "WildChat: 1M ChatGPT Interaction Logs in the Wild", ICLR 2024), LMArena (Arena human-preference 100k and 140k), OpenAssistant (oasst2), the Wikipedia contributors, Hugging Face (FinePDFs-Edu), OpenAI (GSM8K, Cobbe et al. 2021; gpt-oss-120b), Hendrycks et al. (MATH, 2021), NVIDIA (Nemotron-Post-Training-Dataset-v2), Pleias and the AI Alliance (SYNTH), the Qwen team (replies in the seed conversations), and DeepSeek (DeepSeek-V4.1-Flash). Practice problems were written with OpenAI gpt-6.1-sol, Cursor Composer 2.5 and xAI Grok 4.7 (created with Grok). FinePDFs-Edu also contains data from Common Crawl.

ODC-By notice: this dataset contains content from WildChat-4.8M and FinePDFs-Edu, made available under the Open Data Commons Attribution License v1.0.

Licence

other: CC-BY-4.0 by default, with per-row exceptions in the row_licence column. Our rows and our parts of every row (generated answers, templates, prompts we wrote or generated, the data-factory problems) are CC-BY-4.0, Copyright 2026 Udgam Labs LLP. The exceptions:

• Wikipedia rows (597) are CC-BY-SA-4.0 as a whole; if you adapt them, share alike and keep the title and URL.
• WildChat prompts stay ODC-By-1.0, oasst2 prompts Apache-2.0, GSM8K problems MIT, and FinePDFs-Edu metadata ODC-By-1.0 (the documents themselves are not included).
• Arena user prompts, Nemotron and SYNTH first exchanges are CC-BY-4.0 from their publishers.

Laghu's weights are Apache-2.0. Full licence audit and credits: DATA_LICENSES.

Citation

@misc{udgamlabs2026laghusft,
  title  = {Laghu SFT data: the supervised fine-tuning conversations of Laghu},
  author = {Prasad, Jitesh K},
  year   = {2026},
  howpublished = {Udgam Labs},
  url    = {https://huggingface.co/datasets/udgamlabs/laghu-sft}
}

---

About Udgam Labs. Udgam Labs is an independent AI research lab. We work from first principles: before we build, we understand why things work, whether we are training a model, fine-tuning one or applying AI to a new problem. Udgam (उद्गम) is Sanskrit for origin. udgamlabs.com