track png files with git lfs

keep hf version of notebook

add training notebook

add training notebook

add wandb training logs link

add blog and notebook

Expand to 50 tickets with resolution_hints — tickets.py and notebook ALL_TICKETS in sync

Add clarifying comment: loop penalty intentionally omitted in notebook (episode-level concern, live env handles it)

Sync notebook _local_reward: wire resolution_hint + classified_correctly into task3 reward, cls_credit into task2 step-1

Sync notebook Cell 7 graders with graders.py fixes #2 #3 #4 #5 — smoke test passes

Fix #5: accumulate Task 2 classification credit into final score — action scaled to 0.7 max, classify adds up to 0.3, total 1.0

Fix #4: use resolution_hint in reply scoring — category hits 0.03, hint hits 0.05, cap 0.25 (intentional specificity incentive)

Fix #3: track _classified_correctly separately — wrong classification no longer gets free 0.20 credit in Task 3; TODO comment added to Task 2 classify branch

Fix #2: cap _reply_quality at 0.25, add case-insensitive punctuation-stripped matching (weights now sum to exactly 1.0)

Highlight Theme 3.1 + Scaler sub-theme fit, promote GRPO results section

Fix: use raw GitHub URL for chart image to bypass CDN cache

Update: replace broken chart with winning GRPO results (Overall 0.29->0.57)

Fix: add gradio to pyproject.toml deps, update README structure to match actual files

Cleanup: remove junk files, update .gitignore

Remove redundant train_grpo_safe.ipynb

result after no sleep

Fix SFTConfig: move max_seq_length + dataset_text_field to SFTTrainer (trl API change)

Add train_sft.ipynb: SFT pre-training with 1000 gold-label examples before GRPO

Fix AttributeError: 'str' has no .get - add _safe_parse() always returns dict, guard in _local_reward

Fix 500 errors: use LocalEnv for eval (live env is single-instance stateful, breaks under concurrent calls)

Fix sanity checks: use correct seed->ticket mapping, json.dumps for completions

Fix: torch_dtype -> dtype (deprecated warning)

Fix CUDA illegal memory access: drop 4-bit quant, load fp16 natively (0.5B fits in T4), disable DataParallel, fix eval seeds

Sync _local_reward exactly to graders.py: task2 partial credit, task3 efficiency bonus, loop penalty, reply_quality 0-0.5

Expand dataset: 50 tickets x 200 seeds x 3 tasks x 2 steps = ~500 samples, local reward fn, 3 epochs

Fix: single GPU device_map, safe Obs parsing, stricter system prompt (no respond/resolve)

Rewrite: real GRPO using trl.GRPOTrainer with proper KL + clipped ratio + reference model

Fix: separate inference/training modes - use_cache=True for generate, gradient_checkpointing only during train

Kaggle compatibility: auto-detect runtime, fix output paths, remove Colab-only download

Remove Unsloth: use standard HF transformers + PEFT for GRPO training

feat: auto-fallback to local env mirror when live API unreachable

fix: replace remote API with local env for reliable training rewards

fix: support Kaggle secrets via UserSecretsClient

fix: set UNSLOTH_RETURN_LOGITS=1 before model load

fix: set UNSLOTH_RETURN_LOGITS=1 before training loop

fix: patch get_statistics in llama+qwen2 modules directly

fix: monkey-patch unsloth stats check to bypass TimeoutError

fix: remove unsloth, use plain PEFT LoRA - no timeout issues

fix: pin compatible unsloth versions, disable stats timeout

feat: upgrade training notebook to Unsloth (2x faster, 4-bit LoRA)

fix: clean token handling in notebook using Colab Secrets

feat: add GRPO training notebook (token-safe, Colab-ready)

feat: fix scoring, improve prompt, add reward chart

feat: fix scoring formula, improve system prompt, boost reply quality grader - Task1 1.0, Task2 0.60, Task3 0.41, Overall 0.67

fix: update openenv-core version to 0.2.2