--- license: other license_name: brihas-model-licence license_link: LICENSE library_name: gguf pipeline_tag: text-generation tags: - gguf - brihas - cpu - llama.cpp - low-ram language: - en - hi --- # BRIHAS-0.6B BRIHAS-0.6B is the language model behind **BRIHAS**, the built-in assistant of Vayku. It is a small model in GGUF format (Q4_K_M) made to run on an ordinary computer's processor, with no graphics card and little memory. ## What it is - A 0.6 billion-parameter decoder-only language model, quantised to 4 bits (Q4_K_M), in one file: `BRIHAS-0.6B-Q4_K_M.gguf` (397 MB (396,705,280 bytes)). - A derivative of an openly licensed (Apache License 2.0) model. See "Provenance" below for exactly what was and was not changed. - Chat format: the ChatML-style template stored in the file (it supports tool calls through `--jinja`). Thinking can be switched off per request. - Languages: English and Hindi are what BRIHAS is meant for. A model this small is weaker in every language other than English. ## Intended use - The BRIHAS assistant in Vayku: short answers, summaries of the open page, simple browser jobs through tool calls, on **CPU only and low RAM**. - Memory for the engine with the model loaded: about 1.35 GiB (window 4096) to about 1.7 to 1.8 GiB (window 8192). Vayku starts the model only when enough memory is free at that moment (about 1.6 GiB at a window of 4096, about 2 GiB at 8192), so whether a computer with 4 GB of RAM has enough depends on what else is open. - Also fine as a small local model for your own llama.cpp projects where speed and memory matter more than depth. Not intended for: medical, legal or financial advice, anything where a wrong answer causes harm, or any job that needs reliable facts without a check. ## How to run with llama.cpp Vayku downloads this file and starts the engine for you. To run it yourself with [llama.cpp](https://github.com/ggml-org/llama.cpp) (tested with release b12188), these are the arguments the app uses: ``` llama-server --host 127.0.0.1 --port 8080 -m BRIHAS-0.6B-Q4_K_M.gguf --alias brihas -c 8192 -np 1 -t 4 --jinja --no-webui --reasoning-budget 96 --no-reasoning-preserve --reasoning-budget-message "That is enough thinking. I will answer now." -cram 256 -ctxcp 2 ``` Use `-c 4096` on a computer with 8 GB of RAM or less, and `-t` about three quarters of your cores. Then ask a question with thinking off: ``` curl http://127.0.0.1:8080/v1/chat/completions -H "Content-Type: application/json" -d '{"messages":[{"role":"user","content":"What is the capital of France?"}],"max_tokens":64,"temperature":0,"chat_template_kwargs":{"enable_thinking":false}}' ``` The server listens on the loopback address only; keep it that way. Sampling the app uses: temperature 0.65, top_p 0.92, top_k 30, repeat_penalty 1.1, presence_penalty 0.3, frequency_penalty 0.1. ## Size and memory | File | Size | Parameters | Quantisation | Memory with the model loaded | |---|---|---|---|---| | `BRIHAS-0.6B-Q4_K_M.gguf` | 397 MB (396,705,280 bytes) | 0.6 billion | Q4_K_M | about 1.35 GiB (window 4096) to about 1.7 to 1.8 GiB (window 8192) | SHA-256 of the file: `f991263eb7d41244cc0885048d176eafc480fd0604abcbc9466ff15f80ec9417` (also in `SHA256SUMS` and `manifest.json`). ## Limitations - **Small models get facts wrong.** BRIHAS-0.6B will state wrong dates, names, numbers and quotations with complete confidence. Check anything that matters against the page or a source. - It reasons poorly over long chains of steps, and long or tangled pages exceed what it can use well. It manages short, simple jobs best. - Hindi and other non-English output is noticeably weaker than English. - It can repeat itself, ignore an instruction, or call a tool with the wrong argument. The browser guards against some of this; the model itself does not. - It has no knowledge of events after its training data ends, and it has no internet access of its own. - No safety evaluation specific to BRIHAS-0.6B has been done. It may produce biased or inappropriate text. ## Provenance BRIHAS-0.6B is a modified derivative of an openly licensed (Apache License 2.0) language model. It was made by fine-tuning that model with LoRA adapters, merging them into the weights and quantising the result to Q4_K_M. It was **not trained from scratch**. What was trained and how it scored is stated under "Fine-tune details" below; nothing outside that section is claimed. - Packaging: `scripts/finetune/package_release.py` of the Vayku project, version 1.0.0. It rewrites metadata only and compares every tensor of the output with its input; the result is in `manifest.json`. - The name BRIHAS is used by the BRIHAS project. ## Fine-tune details ### What this model is This build is a small fine-tune of the unmodified starting model. The fine-tune was done in three rounds with LoRA (low-rank adapters, about 10 million trainable parameters out of about 600 million). After each round the next one continued from the previous round's adapter. The adapter of the best validation step of the last round (step 14 of 27) was merged into the weights and the result was quantised to Q4_K_M. Nothing was trained from scratch. ### What it was trained for The training aimed at the jobs the BRIHAS assistant does in the browser: - quick answers to general questions, in five answer shapes (plain answer, researcher, thinker, teacher, advisor), kept short and ending cleanly; - follow-up questions that refer to the previous turn ("why?", "and Delhi?", "in Hindi", "simpler"); - grounded answers about the open page, from the lines of the page the browser passes in, including Hinglish and Devanagari pages, quoting the page line instead of guessing, and saying the page does not say it when it does not; - answers from numbered web excerpts, citing the excerpt, preferring the majority when excerpts disagree, and ignoring instructions hidden inside an excerpt; - tool actions (open a page, find, read, click, fill, select, check, scroll, go back, switch tabs) with the arguments in the exact form the browser expects; - a safe hand-back when a field is for a password, a card number or a one-time code (the assistant should leave it to the person instead of filling it); - stopping cleanly: ending after the answer or after a successful action, and saying "too much for me" instead of rambling on a job that is too big. ### What it was trained on Roughly 2,700 training samples were used in total across the three rounds, counting only the steps that were kept (about 1,250 in round 1, 1,300 in round 2 and about 220 in round 3; the same example can appear again in a later round). They come from a pool of about 1,800 distinct compact examples. The examples are made from the project's own code and rules (the browser's command table, its tool formats, its page-line retrieval, its safety rules, its answer-shape rules) plus samples written by a larger open model that were filtered by rules before they were kept. A validation set was kept apart and used to choose the step to keep. The blind test below was written separately and was not used to build or filter any training example. ### How it scored An independent blind test of 178 items was run on this model and, with the same settings, on the unmodified starting model (same quantisation, same engine, the sampling the browser uses, seed 7). Items are scored by fixed rules (accepted and known-wrong answer patterns, tool-call checks), not by opinion. | Category | Items | Starting model right | This model right | |---|---|---|---| | Quick facts | 43 | 21 | 23 | | Answer shapes (lens) | 20 | 15 | 15 | | Follow-ups | 16 | 10 | 12 | | Questions about a page | 34 | 27 | 28 | | Answers from web excerpts | 15 | 12 | 13 | | Tool actions | 15 | 3 | 9 | | Hindi and Hinglish | 15 | 2 | 2 | | Safety | 12 | 9 | 8 | | Stopping cleanly | 8 | 5 | 6 | | **All** | **178** | **104 (58%)** | **116 (65%)** | A second run with another random seed, on the routes that sample (page answers, tool actions, Hindi, safety, stopping; 71 items), gave: tool actions 3 to 10 of 15, page 28 to 31 of 34, safety 7 to 7 of 9, Hindi 0 to 0 of 9, stopping 3 to 1 of 4, all 41 to 49 of 71. Replying in Devanagari when asked to was right in 7 of 8 items (3 of 8 for the starting model) in the first run. The only difference that is large compared with the noise is tool actions (3 to 9 of 15). Every other difference is one or two items out of a small set and should be read as noise, not as an improvement or a loss. ### Limitations, stated plainly - It is a very small model. It still gets many facts wrong (16 of the 43 quick-fact items were answered wrongly) and sometimes states a wrong answer with full confidence. - Hindi is weak: 2 of 15 Hindi and Hinglish items were right, the same as the starting model. It can write Devanagari when asked, but often does not answer the content correctly. - It can still propose filling in a password or a card field. In two safety items of the blind test it did so, where the starting model did not. The browser itself refuses to fill those fields, so the person's secrets are protected by the browser, not by the model. - Safety was not improved overall (9 to 8 of 12) and no claim of safety is made. - For knowledge it does not hold, the assistant uses web lookup in the browser; the model has no internet access on its own and no knowledge of recent events. - The blind test is a small set (178 items, 8 to 43 per category). Confidence intervals are wide (for example 58% to 72% for the whole set), and one or two items either way is within noise. - Checked only with the browser's own prompts and tools. Other prompt formats were not tested. ## Licence Two licences apply, and both files are in this repository: - `LICENSE`: the BRIHAS Model Licence 1.0, which covers what the BRIHAS project added (the fine-tune, the packaging and the names BRIHAS and Vayku). In short: you may use, run, copy and distribute the model, including inside your software and services; you keep the licence files with every copy; you do not use the BRIHAS or Vayku names for a changed version; you do not sell the model file on its own. Read the file for the exact terms. - `LICENSE-APACHE-2.0`: the Apache License 2.0 of the model this one is derived from, with its copyright notice. It continues to apply to the parts that come from that model, and nothing in the BRIHAS terms reduces your rights under it. The `general.license` field inside the GGUF file says `apache-2.0`: that names the licence of the upstream model, which governs those parts; the BRIHAS Model Licence applies in addition. `NOTICE` states that this is a modified derivative and what was changed. The model's outputs come with no warranty.