M37labsorg

Saransh-1.7B

README

License: apache-2.0

Why this model

Length control that actually holds. Asked for one or two sentences, the untuned base model writes 252 words. Saransh writes 24. Length adherence went from 0 percent to 100 percent on short requests and from 0 percent to 91 percent on paragraph requests.
Explicit word budgets. "Summarize in about 120 words" is a trained behaviour, not a hopeful prompt. Roughly a third of the training corpus carries explicit numeric targets.
Faithfulness by construction. Training examples whose reference summary asserted numbers or named entities absent from its own source document were removed, along with every example that attributed a claim to a publication not present in the source.
Runs locally. Quantized GGUF builds work in Ollama, LM Studio and llama.cpp with no GPU required.
Permissive licence. Apache 2.0, inherited from Qwen3, so commercial use is allowed.

Model specifications

Table with columns: Property, Value
Property	Value
Parameters	1,720,574,976 (1.72 B)
Architecture	Qwen3, decoder only transformer
Layers	28
Hidden size	2048
Feed forward size	6144
Attention heads	16 query, 8 key/value (grouped query attention)
Head dimension	128
Activation	SwiGLU
Normalisation	RMSNorm, epsilon 1e-6
Position encoding	Rotary (RoPE)
Tied embeddings	Yes
Vocabulary	151,936 tokens, byte level BPE
Trained context window	8,192 tokens (roughly 12 pages of English prose)
Architectural maximum context	40,960 tokens
Training precision	bfloat16, full parameter fine tune
Objective	Causal language modelling with completion only loss
Chat format	ChatML, no reasoning or thinking block

Length control

Every training example was assigned a length bucket derived from its reference summary, and paired with an instruction matching that bucket. That coupling is what makes the modes distinct behaviours rather than polite suggestions.

Table with columns: Mode, Instruction, Typical output
Mode	Instruction	Typical output
Short	`Summarize the following text in one or two sentences.`	1 to 2 sentences
Medium	`Write a concise summary of the following text in a single paragraph.`	one paragraph
Long	`Write a detailed summary of the following text, covering all the main points.`	multiple paragraphs
Explicit	`Summarize the following text in about 120 words.`

Each bucket was trained with five different phrasings, so paraphrases of these instructions work as well.

Usage

Transformers

python
from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "M37labsorg/Saransh-1.7B"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, dtype="bfloat16", device_map="auto")

text = open("article.txt").read()
msgs = [{"role": "user", "content":
         "Write a concise summary of the following text in a single paragraph."
         "\n\n---\n" + text + "\n---"}]

prompt = tok.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
ids = tok(prompt, return_tensors="pt", add_special_tokens=False).to(model.device)
out = model.generate(**ids, max_new_tokens=420, do_sample=False)
print(tok.decode(out[0][ids.input_ids.shape[1]:], skip_special_tokens=True))

The chat template injects the Saransh system prompt automatically when you do not supply one, so a bare user turn behaves correctly.

Ollama

bash
ollama run hf.co/M37labsorg/Saransh-1.7B-GGUF:Q4_K_M

LM Studio

Search for M37labsorg/Saransh-1.7B-GGUF in the model browser, or place a .gguf file in your local models directory. Set context length to 8192 and temperature to 0.3.

Recommended generation settings

Table with columns: Parameter, Value, Reason
Parameter	Value	Reason
`do_sample`	`False`	Summarization benefits from determinism
`temperature`	0.3 (if sampling)	Low creativity, high fidelity
`repetition_penalty`	1.05	Discourages loops on long inputs
`max_new_tokens`	160 / 420 / 1024

Quantized builds

Available at M37labsorg/Saransh-1.7B-GGUF.

Table with columns: File, Bits, Size, Use case
File	Bits	Size	Use case
`saransh-1.7b-Q4_K_M.gguf`	4	1.1 GB	Default for laptops, best size to quality ratio
`saransh-1.7b-Q5_K_M.gguf`	5	1.3 GB	Slightly higher fidelity
`saransh-1.7b-Q8_0.gguf`	8	1.8 GB	Near lossless

Training corpus

A curated mixture spanning the domains a general summarizer actually encounters, covering news, conversation, legislation, government reporting, literature and scientific writing.

Table with columns: Source, Domain, Contribution
Source	Domain	Contribution
CNN / DailyMail	News articles	Short and medium summaries
XSum	BBC news	Extreme compression, single sentence
Multi-News	Multi document news clusters	Medium and long
DialogSum	Everyday conversation	Short
SAMSum	Messenger style chat	Short
BillSum	United States legislation

The corpus was reduced from 128,787 candidate examples to 97,038 through three successive curation passes.

Data curation

Generic quality filtering removed examples that fail basic summarization criteria: summaries that do not compress the source, implausible compression ratios, truncated references, non prose content, near duplicate documents and anything exceeding the context window. Source documents are never truncated. An over length example is dropped instead, because truncating a document while keeping its full summary teaches the model to invent the missing content.

Faithfulness filtering was added after measuring the reference summaries directly. In the raw corpus, 14.8 percent of numbers and 35.9 percent of proper nouns in XSum reference summaries do not appear in the documents those summaries describe. Multi-News scored 12.0 percent and 22.0 percent on the same measures. References like these teach confabulation, so examples were removed when the summary asserted a number or named entity absent from its own source, with number word normalisation so that "eighteen" in a document supports "18" in its summary.

Attribution filtering removed a specific and common failure. Multi-News reference summaries routinely credit claims to publications that never appear in the source document, phrases of the form "according to a report in the New York Times". This taught the model a template it would then fill with a plausible sounding but invented outlet. Every example carrying an unsupported attribution was dropped, which removed 10,087 Multi-News examples, roughly 46 percent of that dataset.

Length distribution after curation: 39,345 short, 29,241 medium and 28,452 long examples, with 35 percent of all examples carrying an explicit numeric word target.

Evaluation

Measured on 300 held out documents spanning all ten source datasets, greedy decoding, compared against the untuned base model under identical prompts.

Table with columns: Metric, Qwen3-1.7B base, Saransh 1.7B
Metric	Qwen3-1.7B base	Saransh 1.7B
ROUGE-1	0.190	0.387
ROUGE-2	0.049	0.170
ROUGE-L	0.106	0.271

Length adherence

The share of requests whose output lands inside the band that was asked for.

Table with columns: Mode, Base words out, Base on target, Saransh words out, Saransh on target
Mode	Base words out	Base on target	Saransh words out	Saransh on target
Short	252.1	0 percent	23.9	100 percent
Medium	332.4	0 percent	86.9	91 percent
Long	525.2	100 percent	346.7

The base model ignores length instructions entirely. Asked for one or two sentences it produces 252 words. Closing that gap is the central purpose of the fine tune.

ROUGE-L by source

Table with columns: Source, ROUGE-L
Source	ROUGE-L
BillSum	0.396
DialogSum	0.322
SAMSum	0.309
XSum	0.259
PubMed	0.238
Multi-News	0.226
arXiv	0.223
CNN / DailyMail	0.219
GovReport	0.218

Limitations

Every number and proper noun in 300 generated summaries was checked against the document it summarized.

Table with columns: Failure mode, Rate
Failure mode	Rate
Output contains a number absent from the source	3.3 percent
Output contains a name absent from the source	8.3 percent
Output invents a publication name	0.7 percent

The name figure is an upper bound, since it counts morphological variants and legitimate paraphrase as unsupported. The invented publication case is rare but real, inherited from news summarization references, and appears most often on science and news writing. Verify attributions before quoting them.

Further limitations:

English only.
Inputs beyond 8,192 tokens require chunking.
Moderately extractive. About 30 percent of output 8-grams appear verbatim in the source.
Verbose relative to references, overshooting reference length by 20 to 35 percent. Use the explicit word count form when you need a precise budget.
Long mode on a short document converges toward medium mode output, which is correct behaviour, since a short document cannot support a detailed summary.
Summarization can drop or blur detail. Review the output before relying on it for legal, medical or financial decisions.

Intended use

Saransh is intended for summarizing English prose documents: articles, reports, transcripts, papers and correspondence. It is suitable for local and offline deployment, batch document processing and integration into note taking or research tooling.

It is not intended for factual question answering, translation, open ended conversation, or any application where an unverified summary would drive a consequential decision.

Licence and attribution

Released under Apache 2.0, inherited from the base model. Saransh is a derivative work of Qwen/Qwen3-1.7B.

Citation

bibtex
@misc{maithani2026saransh,
  title  = {Saransh: A Length Controlled Small Language Model for Abstractive Summarization},
  author = {Maithani, Aniket for M37Labs},
  year   = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/M37labsorg/Saransh-1.7B}}
}

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.