Math · GSM8K
96.09%
123 / 128 correct
Numeric final-answer match on a fixed 128-case subset.
Bharat 24B by Leo Labs · Released 12 October 2026
Strong math.
Practical Python code.
Meet Bharat 24B, our math and coding focused assistant.
Fine-tuned on 5,000 complete synthetic problem-response pairs, Bharat builds on Mistral Small 3.2 24B Instruct. Try it on your own problems, inspect its reasoning and run its Python solutions.
Bharat's own scorecard
Code · HumanEval
79.69%
51 / 64 passed
Original Python tests on a fixed 64-case subset. Greedy pass@1: one attempt per problem.
These results cover 192 math/code cases from a completed 222-case evaluation run.
Test method and exact scope
The evaluated step 625 adapter was reloaded with its native tokenizer, on an NF4 backbone using original frozen native prompt IDs and greedy decoding. Math used published numeric answers and MathVerify 0.9.0; code used original HumanEval tests in an isolated process. The normalized exact training-question overlap check found zero matches.
The headline math result is a format-tolerant secondary analysis: its checker was chosen after an earlier answer-format artifact and declared before the remaining 100 responses were generated. The registered #### answer-format result is 111/128 (86.72%), retained in the downloadable results. Six math responses reached the generation-token limit; no math/code case was cut by the evaluation wall-time limit. These are fixed subsets, rather than full benchmark or leaderboard results.
5,000 complete responses
2,500 math and 2,500 code examples. One epoch, 625 optimizer updates, 17.5 million assistant target tokens. Complete reasoning and response content, with assistant-only training labels.
A compact download
A 369.7 MB LoRA adapter, native tokenizer and inference helper. The evaluated adapter is preserved byte for byte. Load it with the original Mistral Small 3.2 24B base weights on a CUDA GPU.
Start building with Bharat
Download the repository, install its requirements and use the included native inference helper:
python -m pip install -r requirements.txt
python bharat_inference.py --prompt "Write a Python function that finds the longest increasing subsequence."The helper verifies checkpoint hashes and activates the saved adapter. The model card includes revision pinning and the complete setup.
Base and attribution
Developed by Leo Labs on Mistral Small 3.2 24B Instruct by Mistral AI. Apache 2.0 release; base attribution, license, source-question credits and training details are included in the repository. Bharat is an adapter release; base weights are required.