fix(docker): install musl so prebuilt llama-cpp-python wheel loads on Debian

fix(docker): use prebuilt llama-cpp-python wheel, skip C++ compilation

fix(docker): MAKEFLAGS=-j1 to prevent OOMKilled during llama-cpp-python compile

fix(docker): add build-essential for llama-cpp-python compilation

fix(docker): use prebuilt llama-cpp-python CPU wheel, skip source compile

fix: add blog.md at repo root for hackathon submission URL

feat: wire trained GGUF model (customer-support-grpo-v5) into inference

fix(plots): serve images via GitHub raw URLs, restore results/ for origin

fix(assets): move plot PNGs to frontend/public, remove binary files from results/

fix(docker): remove --prefer-offline from npm ci to install all lockfile packages

feat(frontend): minimal blog redesign — 6 sections, fixed tables, cleaner layout

feat(frontend): white theme, remove carousel, add training plots, fix conversation flow

docs: correct 30→40 steps, add precise Colab metrics to README + blog

fix(readme): update plot URLs to main branch after merge

fix(readme): use GitHub raw URLs for plots (HF Space blocks binary files)

feat: add training plots, update Colab link, embed results in README + blog

blog half done

Merge pull request #14 from lebiraja/feat/blog

the forntend is redesigined and the blog is updated

Merge pull request #13 from lebiraja/feat/refactor

fix(colab): 3 bugs — prompt KeyError, response→respond normalize, max_length warning

feat(colab): v5 notebook — 6-panel dashboard + HF Space + all v5 fixes

feat(colab): v5 notebook — HF Space URL, full parser, FALLBACK-PENALIZED, early stopping

fix(space): serve Next.js frontend via nginx reverse proxy

Merge pull request #12 from lebiraja/feat/refactor

fix(config): update to v5 repo names

fix(train): v5 — JSON preprocessor + dual-token for friend's account

fix(prompt): v4 — add explicit JSON examples for all action types

fix(train): v3 — fix fallback rewards, token budget, steps cap

fix(train): v3 — repair truncated JSON + earlier evals + fix GGUF path

fix(train): v2 — fix collapse root causes + best-ckpt tracking + GGUF export

fix(train): switch to l40sx1 — better availability, 300 steps fits $24 budget

Merge pull request #11 from lebiraja/feat/refactor

fix(train): reduce steps to 250 — guaranteed to complete within $24.05

fix(train): reduce total_steps 400→350 to fit $24.05 budget

feat(train): switch to a100-large, reduce steps/tokens to fit $24 budget

feat(logging): full visibility into every training step

perf(train): optimise for H200 budget — 1000 steps in ~5h for $26

fix(train): also clone log_probs to escape inference_mode context

Merge pull request #10 from lebiraja/feat/refactor

fix(train): clone prompt/completion tensors after inference_mode generation

fix(train): add for_inference/for_training toggles + set lora_dropout=0

fix(train): tokenize-free log-prob recompute + switch default model to Llama 3.1

feat(jobs): default model to Qwen2.5-7B-Instruct, expose TRAIN_MODEL/TOTAL_STEPS

Merge pull request #9 from lebiraja/feat/refactor

fix(train): skip degenerate log-prob recomputes + cap default at 3500 steps

fix(train): allow gradients through compute_log_probs for current policy

fix(train): drop .to(device) calls on bnb 4-bit models — Unsloth places them already

fix(jobs): authenticate git clone with HF_TOKEN — Spaces require auth even when public

fix(training): correct sft_warmstart arg + ensure log dir exists before redirect