# aakaran/finetuning-with-sampling

> Official implementation of Finetuning with Sampling: improving SFT via MCMC sampling of solution traces for chemistry and math tasks. Includes sampling, training, and evaluation scripts.

- Magnitude: 1.3 out of 10 — Steady
- Stars: 74 total · +2 stars measured 2026-10-06, 15:03–17:45 UTC, ≈ 3 by evening
- Star trust: star growth looks organic
- Category: Language models · Language: Python · Created: 2026-10-01 · Last push: 2026-10-05
- GitHub: https://github.com/aakaran/finetuning-with-sampling · Page: https://gitnova.dev/en/r/aakaran/finetuning-with-sampling

## Useful for

- Run projection sampling on chemistry data via sbatch boost_sci.sh
- Train an SFT checkpoint on generated traces with learning_rate=5e-5
- Evaluate trained model accuracy on chemistry and math tasks

## Why it’s here

- Star-counter measurements on 2026-10-06 (UTC), 15:03–17:45: 72 → 74 stars (+2). This is the change over that interval.
- Estimated end-of-day forecast: about +3 stars, using observed gains and the previous day.
- The repository is 5 days old and already has 74 stars. With less than two weeks of history, there's no usual pace to compare the spike against yet.
- Top new repositories this week: #187.

## Star trust

Star growth looks organic.

Star-trust labels are heuristics based on the repository’s behavior, not a check of every stargazer.

## Numbers

- Forks: 9
- Issues and pull requests: 2
- Watchers: 1
- Average over the last week: 12 per day
- Usual pace: too little history (under two weeks)
- Stars in the last hour (measured): 0

## Stars per day, last 10 days (oldest → newest, today is partial)

2026-09-27 … 2026-10-06: 0, 0, 0, 0, 0, 24, 29, 7, 10, 2

## Spotted in now

- Top new repositories this week: #187

## More in this category

1. **Niko1221/Strata** — 7.2 · Breakout · Language models · C++ · +1,501 stars measured 2026-10-06, 00:17–17:34 UTC, ≈ 2,059 by evening
   A C++ local inference engine that runs the 125B MoE model Qwen3.8-Flash-Next on a regular PC with one NVIDIA GPU (12-24 GB) and 64 GB RAM, exposing an OpenAI/Anthropic-compatible API on localhost.
   Full card: https://gitnova.dev/en/r/Niko1221/Strata.md
2. **deepseek-ai/DeepGEMM** — 5.9 · Breakout · Language models · Cuda · +102 stars measured 2026-10-06, 11:46–17:34 UTC, ≈ 142 by evening
   CUDA tensor core kernel library: GEMM (FP8, FP4, BF16), fused MoE, MQA scoring for the indexer. Kernels compile at runtime via DeepJIT, no CUDA build at install time.
   Full card: https://gitnova.dev/en/r/deepseek-ai/DeepGEMM.md
3. **Hiteater-wzm/eeo** — 5.4 · Early signal · Language models · JavaScript · +61 stars measured 2026-10-06, 00:23–17:35 UTC, ≈ 84 by evening
   An open community and toolkit for "Everything Engine Optimization": brands publish machine-readable cards about themselves, and the tooling measures how visible a brand is in AI engine answers.
   Full card: https://gitnova.dev/en/r/Hiteater-wzm/eeo.md
4. **StayLameBro/backburner** — 5.4 · Early signal · Language models · Python · +77 stars measured 2026-10-06, 00:17–17:35 UTC, ≈ 108 by evening
   A llama.cpp fork that plugs an iPhone into a Mac over USB-C and splits the work: the Mac runs layers 1-40, the iPhone runs 41-64 on its GPU, speeding up prefill and allowing context up to 196k-229k tokens. Aimed at people running…
   Full card: https://gitnova.dev/en/r/StayLameBro/backburner.md
5. **empero-org/brewery-ai** — 4.7 · Early signal · Language models · Python · +7 stars measured 2026-10-06, 00:18–17:35 UTC, ≈ 15 by evening
   A console agent that walks users through the full fine-tuning cycle for language and image models via a chat with an AI guide: from picking a base model and data to training on a GPU and publishing to Hugging Face.
   Full card: https://gitnova.dev/en/r/empero-org/brewery-ai.md

---

Magnitude (0–10) measures how fast and how unusually interest in a repository is growing right now. It is not a quality score.
Days are UTC. “So far today” is a fact; “expected by the end of the day” is a forecast.
Summaries and use cases are written by an LLM (DeepSeek V4.1 Flash) from the README and may be inaccurate: verify specific claims (benchmarks, speed, hardware) in the repository itself.
Data as of 2026-10-06 17:51 UTC, updated every 30 minutes.