docs(readme): link raw training logs + training scripts per organizer ask

fix(notebook): rebuild for free CPU Colab — judge-friendly reproducer

docs(readme): align with v2 GRPO run — correct W&B URL, weights, narrative

fix(notebook): install hf_transfer + clear HF_HUB_ENABLE_HF_TRANSFER to fix Colab download crash

fix(notebook): correct reward weights, fix compose_reward_single signature + tuple unpack, add model hint

fix(train): accept wandb_v1_ key format in W&B gate check

fix(grpo): allow FATHOM_USE_VLLM=0 to bypass vLLM rollout and avoid IS-ratio collapse under QLoRA

fix(train): tolerate missing/invalid WANDB_API_KEY, surface key length in smoke

docs(readme): add truth in advertising statement

Update demo space

docs(readme): add demo links, honest fallback, create demo space

feat(space): add HTML index for judge first-impression

fix(reward): align prompt with SFT, soft format, real recursion signal (A.3 + A.4-bis)

feat(train): flip Path 3 -> Path 1 — production 1.5B GRPO on a100-large

fix(grpo): align sys_msg with SFT - emit <answer>...</answer>

docs(D): wire training evidence - link plots from HF model repo, add W&B run

fix(grpo): tail-truncate context inside _to_prompt to satisfy vllm input-length gate

deps: bnb>=0.48.1 for vllm 0.18 4-bit gate

fix(grpo): TRL 1.2 stock path + LoRA double-wrap fix + smoke memory caps

fix(grpo): TRL 1.2 stock path + LoRA double-wrap fix + smoke memory caps

fix(job_train): force CUDA_VISIBLE_DEVICES=0 to dodge TRL multi-GPU entropy bug

fix(model_load): pin 4-bit model to single CUDA device

data: re-upload data/train.jsonl after force-push to main

data: re-upload data/eval.jsonl after force-push to main

data: re-upload data/sft_traces.jsonl after force-push to main

chore(.gitignore): allow data/*.jsonl on main (re-applied after force-push)

fix(sft): TRL 1.2 SFTConfig renamed max_seq_length -> max_length

fix: heredocs use initialize_config_dir absolute path (no caller file = relative breaks)

fix(jobs): bump bitsandbytes to 0.47 (cu128 binary) so GPU dequant works in HF Job runtime image

fix(job_train): drop flash-attn install to avoid build-time OOM (R3 mitigation)

feat(C): demo materials - architecture diagram, demo script, blog draft, viz reward-pie panel, README polish (preflight green)

fix(job_train): pip install huggingface_hub before downloading data files (R1 parity with job_smoke.sh)

config: scripts/job_train.sh \u2014 Path 3 (qwen_0_5b_smoke + GRPO max_steps=50) for sanity-check run before burning 1.5B credits

feat: phase 1 complete — smoke green on HF Jobs, training scripts, plot generator, Colab notebook, submission preflight

clean repo without secrets or data