update: FAL results report with final eval numbers + conclusions

fix: eval crash (ZeroDivisionError) + add eval-only entrypoint (no retraining needed)

add: FAL demo results report (preliminary — eval crash pending fix)

cleanup: remove placeholder files

add: Modal deployment lessons learned — TRL dependency hell postmortem

fix: num_generations=4 batch=4 (TRL 0.28 requires batch divisible by num_generations)

fix: generation_batch_size must be divisible by num_generations — use G=4 batch=4 for T4

fix: remove max_prompt_length (not in TRL 0.28 GRPOConfig), switch to T4 + fp16

fix: install actual weave + pip install llm-blender + mergekit instead of broken stubs

fix: patch TRL source to make vllm import conditional + remove broken vllm stub

fix: vllm stub version 0.0.1 (below TRL's 0.10.2 minimum) so is_vllm_available() returns False

fix: vllm stub needs __version__='0.11.0' so TRL version check doesn't crash

fix: use shell heredoc for stub creation instead of broken f-string interpolation

fix: use heredoc-style stub creation instead of broken one-liner python

fix: drop vllm, patch trl lazy imports to avoid broken optional dep chains

fix: add weave dep (required by trl 0.28 import chain via wandb integration)

fix: switch to trl==0.28.0 (Modal-validated, no broken llm_blender dep chain)

fix: install trl[judges] to pull all transitive deps (mergekit, llm_blender, etc)

fix: complete file with mergekit dep added to pip_install

fix: add mergekit dependency required by trl==0.24.0 import chain

fix: bump transformers>=4.56.1 to match trl==0.24.0 requirement, drop unsloth (use plain PEFT)

fix: resolve dependency conflict — accelerate>=1.4.0 required by trl==0.24.0

add: FAL demo README with Modal run instructions

add: Modal GRPO training app for Future-as-Label calibration demo

add: Future-as-Label Modal implementation — data pipeline, training, evaluation

add: ADR-003 Future-as-Label demo — detailed implementation plan with research validation

add: V4.2 Final Report — complete project retrospective with evidence-based analysis

add: notebook cell insertion script for base vs tuned comparison

add: base vs tuned comparison cell for V4.2 final evaluation

feat(rewards): add sentiment mismatch penalty to prevent extraction reward hacking

fix(probe): use TRL 0.24.0 log keys — rewards/commerce_reward_fn/mean, grad_norm (not train/ prefix)

fix(classifier): reorder _classify_task_type — insights before push to prevent reengajamento misclassification

fix(rewards): 3 bugs from Cell 8 audit — push length/formal, SQL domain, extraction int check

Fix V4.2 audit: show INPUT REVIEW alongside MODEL OUTPUT for proper human scoring

Fix V4.2: task weights 40/40/10/10, full audit completions, interactive input() scoring

Fix V4.2: GDPO + IWU now active in training reward path (not just monitoring)

Add V4.2 GRPO training notebook (Gold Standard, 0.5B)

Create v4_2-handoff.md

docs: add V4.1 run report — detailed evaluation with per-task analysis and V4.2 roadmap

notebooks: add V4.1 GRPO notebook (parser fix, 600 steps, LR 5e-6, constant_with_warmup)

Create v4_1-handoff.md

docs: add V4 run assessment with lessons learned and improvement roadmap

v4: ROOT CAUSE FIX — use standard PEFT not Unsloth get_peft_model (fused LoRA kernels have dtype bug #4891). Revert to load_in_4bit=True, dtype=None matching V3.

v4: fix NF4 fp16/bf16 dtype bug (unsloth #4891) — load_in_4bit=False, 0.5B fits in full bf16 on 24GB

v4: fix fp16/bf16 mismatch — disable Unsloth gradient checkpointing (causes dtype conflict in LoRA QKV kernels at 0.5B)

v4 notebook: fix dtype Half/BFloat16 mismatch (explicit bf16), fix tied embeddings path, fix max_length warning

v4 notebook: fix TypeError crash, suppress warnings, update paths to CWD, add V3 task-aware system prompts

Fix total_mem → total_memory in V4 notebook (PyTorch API)

Add V4 Instruct-Only GRPO notebook implementing ADR-002

ADR-002: V4 Instruct-Only GRPO — revises dual-model plan based on model repo audit