UDGAM LABS · Language model · 152M parameters, long-context base

Laghu-Base-32K

Laghu-Base-32K is Laghu-Base after a long-context stretch: 1,900 more steps (996M tokens) at a 32,768-token context. It finds a hidden code reliably up to 16K tokens and not beyond. These weights (step 1,900) are where supervised fine-tuning of the chat model Laghu started. It is a text completer, not a chat model. Part of a fully open research release.

Try it: the chat model · code on GitHub · udgamlabs.com/models/laghu

Quick start

from transformers import AutoTokenizer, AutoModelForCausalLM

tok = AutoTokenizer.from_pretrained("udgamlabs/Laghu-Base-32K")
model = AutoModelForCausalLM.from_pretrained("udgamlabs/Laghu-Base-32K")
inputs = tok(open("long_document.txt").read() + "\n\nIn summary,", return_tensors="pt")   # adds <bos>
out = model.generate(**inputs, max_new_tokens=80, do_sample=True, temperature=0.8, top_p=0.95)
print(tok.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Use inputs up to 16K tokens. The config allows 32,768 positions, but recall fails beyond about 16K.

Results

Measured by us on 2026-10-19 (scripts in udgam3/runs/ctx32k/ on GitHub).

Hidden-code recall (one 5-digit code in filler text, exact match, 100 trials per length):

On 400 long documents, bits per byte stay flat from 2K to 32K tokens (0.92-0.94); the 2K base rises to 1.39 by 16K.

Cost on short text (same documents and settings for both models):

Where it falls short

• Recall works to 16K tokens and fails completely at 24K and 32K.
• The stretch costs a little short-text quality: −0.76 points on the commonsense average and +1.10% bits per byte on reasoning text, which missed the 1% limit we had set ourselves (we accepted it and say so).
• Only one long-context task (finding a code) was tested; summarising or answering questions over long documents was not measured on this base model.
• Not instruction-tuned.

How it was trained

From the EMA weights of Laghu-Base (step 76,294): 1,900 steps × 524,288 tokens at 32,768 context, on one RTX PRO 6000 (96 GB) for about 4 hours. Data (about 1.0B tokens): 42% long documents (books, papers, government text and code from the Common Pile, already used for our earlier model Gargi), about 8% long modern documents (web, code, Wikipedia), and 50% replay of pretraining cool-down text the main run never read; all 13-gram decontaminated. The data is published as a rebuild script, not re-hosted. The raw training states at steps 1,750 and 1,900 (with optimiser) are in Laghu-Checkpoints; earlier stretch checkpoints were deleted by a storage clean-up on 2026-10-21, our mistake.

Intended use and limitations

A long-context base for research and fine-tuning. No safety tuning; it continues any text, including harmful or false text.

Licence

Apache-2.0. Copyright 2026 Udgam Labs LLP. Training-data credits: DATA_LICENSES.

Citation

@misc{udgamlabs2026laghu,
  title  = {Laghu: a fully open 152M chat model trained from scratch},
  author = {Prasad, Jitesh K},
  year   = {2026},
  howpublished = {Udgam Labs},
  url    = {https://huggingface.co/udgamlabs/Laghu}
}

---

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