nraptisss/tmf921-intent-training

Add nohup launcher for focused 4-layer training

Add focused 4-layer SFT script: train only on tmf921/camara/3gpp/etsi_zsm (removes weak O1/A1/lifecycle layers)

Fix RFT: batch generation, reduce to 200 prompts x 8 samples (~24h feasible on RTX 6000 Ada)

Add Best-of-N rejection sampling + RFT pipeline for value fidelity improvement

Confirm GRPO negative result with correct evaluation (SFT-merged + GRPO adapter). Result validated.

Add correct GRPO evaluation script that loads SFT-merged base + GRPO adapter

Add IEEE ICC/Globecom paper draft (IEEEtran format) with full references and TikZ figures

Add GRPO experiment journal entries (v1, v2, v3) — negative results documented with root cause analysis

GRPO v3: fix truncation — use full 1536 completion length + G=4, we have 42GB headroom

Fix GRPO v2: lower temp=0.3, dense reward shaping, higher beta=0.1 to stay near SFT, G=2 safe for 48GB

Tune GRPO hyperparams for RTX 6000 Ada 48GB: reduce steps/grad_accum/completion_length for ~6h runtime, prevent OOM

Fix GRPO script: skip CPU merge, load base+adapter directly on GPU in 4-bit to avoid RAM OOM

Add GRPO nohup launch script for RTX 6000 Ada / A100

Add GRPO post-SFT training script with multi-component reward (JSON validity + key F1 + value F1 + layer-weighted bonus)

Upload LEAKAGE_ANALYSIS.md

Upload scripts/run_all_baselines.sh

Add baseline evaluation script for Llama/GPT-4o-mini comparison

Upload paper/main.bbl with huggingface_hub

Upload paper/main.pdf with huggingface_hub

Upload paper/main.tex with huggingface_hub

Update paper/main.tex with semantic evaluator results

Update paper/tables.md with semantic evaluator results

Update PROJECT_JOURNAL.md with semantic evaluator results

Update results/semantic/o1_a1_stage1_vs_stage2_summary.json with semantic evaluator results

Add O1/A1 semantic comparison artifact

Fix semantic evaluator to recover metadata by prediction id

Add prototype semantic evaluator for O1 NRM and A1 policy

Update ML Intern artifact metadata

Restore full IEEE LaTeX main.tex content

Upload paper/main.tex

Add BibTeX references for IEEE paper

Upload paper/main.tex

Add first manuscript draft

Add zero-shot baseline summary JSON

Update paper tables with zero-shot baseline

Add zero-shot vs fine-tuned baseline summary

Restore and update project journal with zero-shot baseline

Upload PROJECT_JOURNAL.md

Upload paper/tables.md

Upload folder using huggingface_hub

Add sampled zero-shot baseline runner

Add stage1 evaluation reproduction script

Add reproducibility checklist

Upload paper/tables.md

Upload paper/outline.md

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Restore README with results packaging instructions

Upload README.md

Add qualitative failure example sampler