# AGENT.md

This document is a working guide for future Codex sessions on the ABForge RL project.

It focuses on:

- local repo structure
- the Misha cluster workflow
- current Task 1 / Task 2 training setup
- API judge vs local judge usage
- checkpoint / resume behavior
- inference workflow
- common pitfalls

This file reflects the current project state as of April 2026.

## 1. Project Overview

This repository is used for RL training on scientific ablation-design tasks.

There are two tasks:

- Task 1: identify key ablation objectives / research questions
- Task 2: produce a rigorous ablation experiment plan

Main local directories:

- `verl_proj/`: customized `verl` training code
- `reward_part/`: local reward servers and judge services
- `readme.md`: project overview

Important tracked scripts in this repo:

- Task 1 preprocess:
  - `/Users/xucai/Project/verl_with_llm-as-judge-master/verl_proj/examples/data_preprocess/abforge_task1_rl.py`
- Task 2 preprocess:
  - `/Users/xucai/Project/verl_with_llm-as-judge-master/verl_proj/examples/data_preprocess/abforge_task2_rl.py`
- Task 1 local judge reward:
  - `/Users/xucai/Project/verl_with_llm-as-judge-master/reward_part/task1_call_api.py`
  - `/Users/xucai/Project/verl_with_llm-as-judge-master/reward_part/task1_rw.py`
- Task 2 local judge reward:
  - `/Users/xucai/Project/verl_with_llm-as-judge-master/reward_part/task2_call_api.py`
  - `/Users/xucai/Project/verl_with_llm-as-judge-master/reward_part/task2_rw.py`
- Task 1 Azure/API judge reward:
  - `/Users/xucai/Project/verl_with_llm-as-judge-master/reward_part/azure_task1_judge/task1_azure_reward.py`
- Task 2 Azure/API judge reward:
  - `/Users/xucai/Project/verl_with_llm-as-judge-master/reward_part/azure_task2_judge/task2_azure_reward.py`

Tracked launchers:

- `/Users/xucai/Project/verl_with_llm-as-judge-master/verl_proj/examples/grpo_trainer/run_abforge_qwen3_8b_task1_hf.sh`
- `/Users/xucai/Project/verl_with_llm-as-judge-master/verl_proj/examples/grpo_trainer/run_abforge_qwen3_8b_task1_vllm.sh`
- `/Users/xucai/Project/verl_with_llm-as-judge-master/verl_proj/examples/grpo_trainer/run_abforge_qwen3_8b_task1_vllm_azurejudge.sh`
- `/Users/xucai/Project/verl_with_llm-as-judge-master/verl_proj/examples/grpo_trainer/run_abforge_qwen3_8b_hf.sh`
- `/Users/xucai/Project/verl_with_llm-as-judge-master/verl_proj/examples/grpo_trainer/run_abforge_qwen3_8b_vllm.sh`
- `/Users/xucai/Project/verl_with_llm-as-judge-master/verl_proj/examples/grpo_trainer/run_abforge_qwen3_8b_vllm_azurejudge.sh`
- `/Users/xucai/Project/verl_with_llm-as-judge-master/verl_proj/examples/grpo_trainer/run_abforge_qwen3_8b_vllm_qwen3_32b_judge.sh`

## 2. Misha Path Map

The user works on Yale Misha.

There are three path classes that matter:

### A. Home

- `/gpfs/radev/home/yz979`

Current role:

- shell login location
- not used for training artifacts
- do not write caches, logs, checkpoints, or model downloads here

### B. Project

- `/gpfs/radev/project/cohan/yz979/xucai/Abforge_Training`

Recommended variable:

```bash
export CODE_ROOT=/gpfs/radev/project/cohan/yz979/xucai/Abforge_Training
```

This is the canonical code / repo root on Misha.

What lives here now:

- `.git/`
- `readme.md`
- `reward_part/`
- `verl_proj/`
- `slurm/`
- small `data/` directory inside repo root
- reference PDF

What should be treated as living under `project`:

- Git-tracked code
- operational Misha slurm files in `$CODE_ROOT/slurm`
- the `verl_proj` code tree
- local reward service code in `$CODE_ROOT/reward_part`

Important real paths currently present under `project`:

- root slurm:
  - `/gpfs/radev/project/cohan/yz979/xucai/Abforge_Training/slurm`
- reward code:
  - `/gpfs/radev/project/cohan/yz979/xucai/Abforge_Training/reward_part`
- verl code:
  - `/gpfs/radev/project/cohan/yz979/xucai/Abforge_Training/verl_proj`
- model merger script:
  - `/gpfs/radev/project/cohan/yz979/xucai/Abforge_Training/verl_proj/scripts/model_merger.py`

### C. Scratch

- `/gpfs/radev/scratch/cohan/yz979/xucai/Abforge_Training`

Recommended variable:

```bash
export WORK_ROOT=/gpfs/radev/scratch/cohan/yz979/xucai/Abforge_Training
```

This is the canonical runtime / artifact root on Misha.

What lives here now:

- `.cache/`
- `.config/`
- `conda/`
- `data/`
- `hf_cache/`
- `infer/`
- `logs/`
- `models/`
- `pip_cache/`
- `secrets/`
- `tmp/`
- `checkpoints/`
- `run_inference_local.py`

What should be treated as living under `scratch`:

- conda envs
- model downloads
- parquet data
- raw JSONL data
- checkpoints
- logs
- inference outputs
- secrets
- temp files
- caches

Important real paths currently present under `scratch`:

- inference script:
  - `/gpfs/radev/scratch/cohan/yz979/xucai/Abforge_Training/run_inference_local.py`
- conda envs:
  - `/gpfs/radev/scratch/cohan/yz979/xucai/Abforge_Training/conda/abforge`
  - `/gpfs/radev/scratch/cohan/yz979/xucai/Abforge_Training/conda/abforge_vllm`
- models:
  - `/gpfs/radev/scratch/cohan/yz979/xucai/Abforge_Training/models/Qwen3-8B`
  - `/gpfs/radev/scratch/cohan/yz979/xucai/Abforge_Training/models/Qwen3-32B`
- secrets:
  - `/gpfs/radev/scratch/cohan/yz979/xucai/Abforge_Training/secrets/azure_openai_task1.env`
  - `/gpfs/radev/scratch/cohan/yz979/xucai/Abforge_Training/secrets/azure_openai_task2.env`

### Rule of Thumb

Use:

- `project` for code and slurm
- `scratch` for everything large or runtime-generated
- `home` only as login shell location

Do not assume a script in local repo also exists on Misha in the same place.
Example:

- local repo inference script may be under `scripts/`
- actual Misha inference script currently lives at:
  - `$WORK_ROOT/run_inference_local.py`

## 3. What Exists Only On Server

Not everything is tracked in Git.

Important server-only or server-maintained items:

- root `slurm/` directory under `$CODE_ROOT`
  - this contains the actual Misha submission scripts used in practice
  - do not assume these are fully tracked in Git
- `/gpfs/radev/scratch/cohan/yz979/xucai/Abforge_Training/run_inference_local.py`
  - this script has been used for local inference on Misha
  - it may exist only in scratch, not in Git
- secrets files under `$WORK_ROOT/secrets`
  - never commit secrets

When helping the user, explicitly distinguish:

- tracked code in the repo
- server-only operational files

## 4. Current Misha File Inventory

These are the currently confirmed server paths and should be treated as authoritative unless the user says they changed them.

### Root slurm files on Misha

Current files under `$CODE_ROOT/slurm`:

- `/gpfs/radev/project/cohan/yz979/xucai/Abforge_Training/slurm/task1_azure_reward.slurm`
- `/gpfs/radev/project/cohan/yz979/xucai/Abforge_Training/slurm/task1_judge_qwen3_32b.slurm`
- `/gpfs/radev/project/cohan/yz979/xucai/Abforge_Training/slurm/task1_train_vllm_azurejudge.slurm`
- `/gpfs/radev/project/cohan/yz979/xucai/Abforge_Training/slurm/task1_train_vllm_qwen3_32b.slurm`
- `/gpfs/radev/project/cohan/yz979/xucai/Abforge_Training/slurm/task2_azure_reward.slurm`
- `/gpfs/radev/project/cohan/yz979/xucai/Abforge_Training/slurm/task2_judge_qwen3_32b.slurm`
- `/gpfs/radev/project/cohan/yz979/xucai/Abforge_Training/slurm/task2_judge.slurm`
- `/gpfs/radev/project/cohan/yz979/xucai/Abforge_Training/slurm/task2_train.slurm`
- `/gpfs/radev/project/cohan/yz979/xucai/Abforge_Training/slurm/task2_train_vllm_azurejudge.slurm`
- `/gpfs/radev/project/cohan/yz979/xucai/Abforge_Training/slurm/task2_train_vllm_qwen3_32b.slurm`
- `/gpfs/radev/project/cohan/yz979/xucai/Abforge_Training/slurm/task2_train_vllm.slurm`

### Scratch data files

Current files under `$WORK_ROOT/data`:

- `/gpfs/radev/scratch/cohan/yz979/xucai/Abforge_Training/data/abforge_task1_rl_full/train.parquet`
- `/gpfs/radev/scratch/cohan/yz979/xucai/Abforge_Training/data/abforge_task1_rl_full/val.parquet`
- `/gpfs/radev/scratch/cohan/yz979/xucai/Abforge_Training/data/abforge_task2_rl_full/train.parquet`
- `/gpfs/radev/scratch/cohan/yz979/xucai/Abforge_Training/data/abforge_task2_rl_full/val.parquet`
- `/gpfs/radev/scratch/cohan/yz979/xucai/Abforge_Training/data/bench_data_4_subset50_simple.jsonl`
- `/gpfs/radev/scratch/cohan/yz979/xucai/Abforge_Training/data/rl_30000.jsonl`
- `/gpfs/radev/scratch/cohan/yz979/xucai/Abforge_Training/data/task1/train.parquet`
- `/gpfs/radev/scratch/cohan/yz979/xucai/Abforge_Training/data/task1/val.parquet`
- `/gpfs/radev/scratch/cohan/yz979/xucai/Abforge_Training/data/task2/train.parquet`
- `/gpfs/radev/scratch/cohan/yz979/xucai/Abforge_Training/data/task2/val.parquet`

### Scratch inference files

Confirmed inference outputs under `$WORK_ROOT/infer`:

- `/gpfs/radev/scratch/cohan/yz979/xucai/Abforge_Training/infer/base_qwen3_infer_task2.jsonl`
- `/gpfs/radev/scratch/cohan/yz979/xucai/Abforge_Training/infer/ckpt28_qwen3_infer_task2.jsonl`
- `/gpfs/radev/scratch/cohan/yz979/xucai/Abforge_Training/infer/task1_api_ckpt14_merged_hf`
- `/gpfs/radev/scratch/cohan/yz979/xucai/Abforge_Training/infer/task1_api_ckpt14_qwen3_infer_task1.jsonl`
- `/gpfs/radev/scratch/cohan/yz979/xucai/Abforge_Training/infer/task1_base_qwen3_infer_task1.jsonl`
- `/gpfs/radev/scratch/cohan/yz979/xucai/Abforge_Training/infer/task2_api_ckpt14_merged_hf`
- `/gpfs/radev/scratch/cohan/yz979/xucai/Abforge_Training/infer/task2_api_ckpt14_qwen3_infer_task2.jsonl`

### Checkpoint roots currently in use

Scratch-side checkpoint roots:

- `/gpfs/radev/scratch/cohan/yz979/xucai/Abforge_Training/checkpoints/abforge_task2/qwen3_8b_grpo_vllm_azurejudge`
- `/gpfs/radev/scratch/cohan/yz979/xucai/Abforge_Training/checkpoints/abforge_task2/qwen3_8b_grpo_vllm_azurejudge_full`

Project-side historical checkpoint roots:

- `/gpfs/radev/project/cohan/yz979/xucai/Abforge_Training/verl_proj/checkpoints/abforge_task1/qwen3_8b_grpo_task1_vllm_azurejudge`
- `/gpfs/radev/project/cohan/yz979/xucai/Abforge_Training/verl_proj/checkpoints/abforge_task2/qwen3_8b_grpo_vllm`

Interpretation:

- older experiments may have written checkpoints under `project/verl_proj/checkpoints`
- new work should prefer `scratch/checkpoints`

## 5. Current Preferred Training Strategy

Current decision:

- API judge with `gpt-5.4-mini` is preferred for RL training
- local Qwen3-32B judge exists, but is much slower
- Task 1 and Task 2 are trained as separate specialist models
- Task 2 should not continue from Task 1 checkpoint

So the intended setup is:

- Task 1 formal RL:
  - base Qwen3-8B actor
  - Task 1 API judge
  - full Task 1 parquet
- Task 2 formal RL:
  - base Qwen3-8B actor
  - Task 2 API judge
  - full Task 2 parquet

## 6. API Judge Configuration

Task 1 Azure/API judge:

- service code:
  - `/Users/xucai/Project/verl_with_llm-as-judge-master/reward_part/azure_task1_judge/task1_azure_reward.py`
- server secret:
  - `$WORK_ROOT/secrets/azure_openai_task1.env`
- default port:
  - `6012`
- endpoint path:
  - `/get_reward_task1`

Task 2 Azure/API judge:

- service code:
  - `/Users/xucai/Project/verl_with_llm-as-judge-master/reward_part/azure_task2_judge/task2_azure_reward.py`
- server secret:
  - `$WORK_ROOT/secrets/azure_openai_task2.env`
- default port:
  - `6011`
- endpoint path:
  - `/get_reward_task2`

Expected env file contents:

```bash
export AZURE_OPENAI_ENDPOINT="https://.../openai/responses?api-version=2025-04-01-preview"
export AZURE_OPENAI_MODEL="gpt-5.4-mini"
export AZURE_OPENAI_API_KEY="..."
```

Recommended Task 2 API robustness settings:

```bash
export TASK2_AZURE_REQUEST_TIMEOUT=300
export TASK2_AZURE_MAX_RETRIES=5
export TASK2_AZURE_MAX_OUTPUT_TOKENS=1024
```

Equivalent Task 1 settings may be added with `TASK1_...` names.

Never write the actual API key into Git-tracked files.

## 7. Conda Environments

Confirmed envs on Misha:

- `/gpfs/radev/scratch/cohan/yz979/xucai/Abforge_Training/conda/abforge`
- `/gpfs/radev/scratch/cohan/yz979/xucai/Abforge_Training/conda/abforge_vllm`

Recommended usage:

- use `abforge_vllm` for:
  - vLLM rollout training
  - API reward services
  - model merger
  - local inference
- use `abforge` only if the user explicitly refers to an older local-judge workflow that depends on it

Typical activation:

```bash
module reset
module load miniconda
module load GCCcore/13.3.0
module load CUDA/12.4.1
source $(conda info --base)/etc/profile.d/conda.sh
conda activate /gpfs/radev/scratch/cohan/yz979/xucai/Abforge_Training/conda/abforge_vllm
```

## 8. Data and Preprocessing

Raw full RL data was uploaded as shared JSONL and then separately preprocessed for Task 1 and Task 2.

Task 1 preprocess:

- `/Users/xucai/Project/verl_with_llm-as-judge-master/verl_proj/examples/data_preprocess/abforge_task1_rl.py`

Task 2 preprocess:

- `/Users/xucai/Project/verl_with_llm-as-judge-master/verl_proj/examples/data_preprocess/abforge_task2_rl.py`

Important note:

- Task 2 `test1000` and `full` use the same field mapping and prompt template
- fields used in Task 2:
  - `Content`
  - `Goal`
  - `Rubric`
  - `refined_standard_plan`

Small test parquet:

- Task 1:
  - `$WORK_ROOT/data/task1/train.parquet`
  - `$WORK_ROOT/data/task1/val.parquet`
- Task 2:
  - `$WORK_ROOT/data/task2/train.parquet`
  - `$WORK_ROOT/data/task2/val.parquet`

Formal parquet:

- Task 1:
  - `$WORK_ROOT/data/abforge_task1_rl_full/train.parquet`
  - `$WORK_ROOT/data/abforge_task1_rl_full/val.parquet`
- Task 2:
  - `$WORK_ROOT/data/abforge_task2_rl_full/train.parquet`
  - `$WORK_ROOT/data/abforge_task2_rl_full/val.parquet`

## 9. Formal API Training Settings

For formal API training on the 30k dataset, prefer:

- `train_batch_size=64`
- `trainer.total_epochs=1`
- `trainer.save_freq=100`
- checkpoints written to scratch
- experiment name suffixed with `_full`

Why:

- with ~29k train rows and batch size 64, one epoch is about 453 steps
- `save_freq=100` gives about 5 saves per epoch
- one epoch is much easier to fit in a 48h wall-clock budget

Recommended Task 2 formal API launcher override pattern:

```bash
bash examples/grpo_trainer/run_abforge_qwen3_8b_vllm_azurejudge.sh \
  trainer.save_freq=100 \
  trainer.total_epochs=1 \
  trainer.default_local_dir=$WORK_ROOT/checkpoints/abforge_task2/qwen3_8b_grpo_vllm_azurejudge_full \
  trainer.experiment_name=qwen3_8b_grpo_vllm_azurejudge_full
```

Recommended Task 1 formal API launcher override pattern:

```bash
bash examples/grpo_trainer/run_abforge_qwen3_8b_task1_vllm_azurejudge.sh \
  trainer.save_freq=100 \
  trainer.total_epochs=1 \
  trainer.default_local_dir=$WORK_ROOT/checkpoints/abforge_task1/qwen3_8b_grpo_task1_vllm_azurejudge_full \
  trainer.experiment_name=qwen3_8b_grpo_task1_vllm_azurejudge_full
```

## 10. Resume Behavior

This `verl` setup supports automatic resume.

Relevant code:

- `/Users/xucai/Project/verl_with_llm-as-judge-master/verl_proj/verl/trainer/ppo/ray_trainer.py`
- `resume_mode` default in:
  - `/Users/xucai/Project/verl_with_llm-as-judge-master/verl_proj/verl/trainer/config/ppo_trainer.yaml`

Behavior:

- checkpoints are saved under `trainer.default_local_dir/global_step_xxx`
- latest step is recorded in:
  - `latest_checkpointed_iteration.txt`
- if `resume_mode=auto` and the directory exists, training resumes automatically

Recommended workflow for 2 epochs:

1. first job:
   - `trainer.total_epochs=1`
2. second job:
   - keep the same `trainer.default_local_dir`
   - change to `trainer.total_epochs=2`
3. submit again

The second job should resume from the latest checkpoint and continue.

## 11. Slurm Usage on Misha

Many real submission scripts live only on the server under:

- `$CODE_ROOT/slurm`

Common files that have existed there:

- `task1_azure_reward.slurm`
- `task2_azure_reward.slurm`
- `task1_train_vllm_azurejudge.slurm`
- `task2_train_vllm_azurejudge.slurm`
- `task1_judge_qwen3_32b.slurm`
- `task2_judge_qwen3_32b.slurm`
- `task1_train_vllm_qwen3_32b.slurm`
- `task2_train_vllm_qwen3_32b.slurm`

API reward services should be started first, then the printed `REWARD_URL` should be copied into the corresponding train slurm.

Typical operational order:

1. `sbatch slurm/task2_azure_reward.slurm`
2. read log under `$WORK_ROOT/logs/task2_azure_reward_<JOBID>.out`
3. update `export REWARD_URL=...` in train slurm
4. `sbatch slurm/task2_train_vllm_azurejudge.slurm`

Same pattern applies to Task 1.

## 12. Checkpoint Locations Used So Far

Examples from previous runs:

Task 1 API test:

- checkpoint dir:
  - `/gpfs/radev/project/cohan/yz979/xucai/Abforge_Training/verl_proj/checkpoints/abforge_task1/qwen3_8b_grpo_task1_vllm_azurejudge/global_step_14`

Task 2 API test:

- checkpoint dir:
  - `/gpfs/radev/scratch/cohan/yz979/xucai/Abforge_Training/checkpoints/abforge_task2/qwen3_8b_grpo_vllm_azurejudge/global_step_14`

Formal Task 2 API run was configured to use:

- `/gpfs/radev/scratch/cohan/yz979/xucai/Abforge_Training/checkpoints/abforge_task2/qwen3_8b_grpo_vllm_azurejudge_full`

Preferred rule:

- always use scratch for new checkpoints

## 13. Inference Workflow

There are two inference cases.

### A. Base model inference

Use the base HF model directly.

- model path:
  - `$WORK_ROOT/models/Qwen3-8B`

Example Task 1 base inference:

```bash
python3 $WORK_ROOT/run_inference_local.py \
  --task 1 \
  --input $WORK_ROOT/data/bench_data_4_subset50_simple.jsonl \
  --output-prefix $WORK_ROOT/infer/task1_base_qwen3 \
  --model-path $WORK_ROOT/models/Qwen3-8B \
  --tokenizer-path $WORK_ROOT/models/Qwen3-8B \
  --trust-remote-code \
  --dtype bf16 \
  --device-map auto \
  --max-new-tokens 1024 \
  --overwrite
```

Important real Misha inference script path:

- `/gpfs/radev/scratch/cohan/yz979/xucai/Abforge_Training/run_inference_local.py`

Do not assume the Git-tracked local script path exists on Misha.

### B. RL checkpoint inference

FSDP actor checkpoints must first be merged into HF format.

Model merger:

- `/Users/xucai/Project/verl_with_llm-as-judge-master/verl_proj/scripts/model_merger.py`

Important:

- `--local_dir` must point to `global_step_xxx/actor`, not the full checkpoint root

Example:

```bash
python3 $CODE_ROOT/verl_proj/scripts/model_merger.py \
  --backend fsdp \
  --hf_model_path $WORK_ROOT/models/Qwen3-8B \
  --local_dir $WORK_ROOT/checkpoints/.../global_step_14/actor \
  --target_dir $WORK_ROOT/infer/task2_api_ckpt14_merged_hf
```

Then run inference with the merged HF directory:

```bash
python3 $WORK_ROOT/run_inference_local.py \
  --task 2 \
  --input $WORK_ROOT/data/bench_data_4_subset50_simple.jsonl \
  --output-prefix $WORK_ROOT/infer/task2_api_ckpt14_qwen3 \
  --model-path $WORK_ROOT/infer/task2_api_ckpt14_merged_hf \
  --tokenizer-path $WORK_ROOT/models/Qwen3-8B \
  --trust-remote-code \
  --dtype bf16 \
  --device-map auto \
  --max-new-tokens 2048 \
  --overwrite
```

## 14. Bench Inference Data

Known inference input on Misha:

- `$WORK_ROOT/data/bench_data_4_subset50_simple.jsonl`

This has been used successfully for:

- base model inference
- Task 1 API checkpoint inference
- Task 2 API checkpoint inference

## 15. Interactive GPU Sessions

For checkpoint merge + local inference, use a single-GPU interactive allocation.

Typical command:

```bash
salloc -p gpu --gpus=1 --cpus-per-task=8 --mem=64G -t 08:00:00
```

Then initialize:

```bash
module reset
module load miniconda
module load GCCcore/13.3.0
module load CUDA/12.4.1
source $(conda info --base)/etc/profile.d/conda.sh
conda activate $WORK_ROOT/conda/abforge_vllm
```

## 16. Git State and Server Mutability

The Misha repo can be dirty.

At last inspection, `git status --short` on `$CODE_ROOT` showed:

- modified tracked files
- untracked `slurm/`
- untracked historical outputs and checkpoints

Interpretation for future Codex sessions:

- do not assume server repo is clean
- do not `git reset --hard`
- do not delete `slurm/` or `verl_proj/checkpoints/` blindly
- if editing tracked code, inspect local diffs first
- if editing server-only slurm files, treat them as operational config, not as Git-tracked source of truth

Git remote currently is:

- `git@github.com:SlowGuess/Abforge_Training.git`

Current branch on Misha:

- `main`

## 17. Common Pitfalls

### 1. API reward job is down

Symptom:

- `curl http://<node>:6011/health` fails
- train job will not get rewards

Fix:

- resubmit the reward slurm
- update `REWARD_URL` in train slurm

### 2. Wrong `REWARD_URL`

Each reward service starts on a new node and prints a new URL.
Do not reuse old node names blindly.

### 3. Checkpoint save failures

Observed before on Task 2 API test:

- writes failed during checkpoint save
- fix was to:
  - save to scratch
  - reduce save frequency

### 4. Confusing test vs formal lines

Keep these separate:

- test:
  - small parquet in `$WORK_ROOT/data/task1` or `$WORK_ROOT/data/task2`
- formal:
  - full parquet in `$WORK_ROOT/data/abforge_task1_rl_full` and `$WORK_ROOT/data/abforge_task2_rl_full`

Use `_full` in the experiment name for formal runs.

### 5. `run_inference_local.py` location confusion

Do not assume it lives in Git.

On Misha it has been used from:

- `$WORK_ROOT/run_inference_local.py`

### 6. `project` vs `scratch` checkpoint confusion

Historical checkpoints exist in both places.

When asked to merge or infer from a checkpoint, always verify whether the checkpoint lives under:

- `$WORK_ROOT/checkpoints/...`
or
- `$CODE_ROOT/verl_proj/checkpoints/...`

Do not assume all checkpoints are on scratch.

### 7. Prompt filtering log

`Filtering prompts longer than 4096 tokens` is a preprocessing / dataset check stage.
It does not automatically mean something is broken.

For Task 2 formal run, 29k train rows with batch size 64 giving about `453` steps is normal and consistent with floor-style batching.

## 18. Local Judge Status

Local judge was explored but is not currently preferred for formal training.

Reasons:

- response is much slower than API judge
- GPU jobs showed very high idle time
- bottleneck is mostly reward throughput and orchestration

Still, local judge code exists and may be useful for future experimentation:

- Task 1 local judge:
  - `/Users/xucai/Project/verl_with_llm-as-judge-master/reward_part/task1_call_api.py`
- Task 2 local judge:
  - `/Users/xucai/Project/verl_with_llm-as-judge-master/reward_part/task2_call_api.py`
- faster Task 2 local judge launcher:
  - `/Users/xucai/Project/verl_with_llm-as-judge-master/verl_proj/examples/grpo_trainer/run_abforge_qwen3_8b_vllm_qwen3_32b_judge.sh`

## 19. Recommended Default Actions for Future Codex Sessions

When assisting on this repo:

1. Ask whether the user wants:
   - test line
   - formal line
   - Task 1
   - Task 2
   - API judge
   - local judge
2. Assume API judge is the preferred formal path unless the user says otherwise.
3. Use scratch for:
   - checkpoints
   - caches
   - secrets
   - inference outputs
4. Treat root `slurm/` files on Misha as operational state, even if not tracked in Git.
5. Before launching training, verify:
   - reward service is alive
   - `REWARD_URL` is current
   - data path is correct
   - experiment name is correct
   - checkpoint path is on scratch
6. Before resuming training, verify whether the target checkpoint directory already exists.

## 20. Minimal Operational Checklist

Before training:

1. set:
   - `CODE_ROOT`
   - `WORK_ROOT`
2. confirm correct conda env
3. confirm reward service is up
4. confirm `REWARD_URL` in slurm is current
5. confirm data path points to test or full dataset intentionally
6. confirm checkpoint root is on scratch for new runs
7. confirm experiment name matches test vs formal intent

Before inference:

1. verify whether base model or merged RL checkpoint is needed
2. if RL checkpoint:
   - identify exact checkpoint root
   - identify whether it is under `project` or `scratch`
   - merge `global_step_xxx/actor` with `model_merger.py`
3. use:
   - `$WORK_ROOT/run_inference_local.py`
4. use:
   - `$WORK_ROOT/data/bench_data_4_subset50_simple.jsonl`
   unless the user specifies another input

## 21. Current Known Good Runs

These are workflows that have already been exercised successfully and should be treated as the safest starting points.

### A. Task 1 API judge test training

Status:

- completed successfully

Observed job:

- training job `1502537`
- reward job `1502526`

Reward URL used during successful run:

- `http://r817u15n02.misha.ycrc.yale.edu:6012/get_reward_task1`

Checkpoint produced:

- `/gpfs/radev/project/cohan/yz979/xucai/Abforge_Training/verl_proj/checkpoints/abforge_task1/qwen3_8b_grpo_task1_vllm_azurejudge/global_step_14`

Meaning:

- Task 1 API reward service worked
- Task 1 RL loop worked
- final checkpoint was saved

### B. Task 2 API judge test training

Status:

- completed successfully after moving checkpoint writing to scratch and reducing save frequency pressure

Observed successful job:

- training job `1504216`

Reward service job seen before successful test:

- reward job `1504288`

Checkpoint produced:

- `/gpfs/radev/scratch/cohan/yz979/xucai/Abforge_Training/checkpoints/abforge_task2/qwen3_8b_grpo_vllm_azurejudge/global_step_14`

Meaning:

- Task 2 API reward service worked
- Task 2 RL loop worked
- final checkpoint was saved to scratch

### C. Task 2 formal API training

Status:

- formal run started successfully on full data

Observed job:

- training job `1504289`

Observed training progress:

- `Training Progress: 0/453`

Interpretation:

- full train split size is about 29000
- batch size 64 gives 453 steps via floor-style batching
- this is normal and not by itself evidence of abnormal filtering

Formal checkpoint path in use:

- `/gpfs/radev/scratch/cohan/yz979/xucai/Abforge_Training/checkpoints/abforge_task2/qwen3_8b_grpo_vllm_azurejudge_full`

Observed saved steps:

- `global_step_100`
- `global_step_200`

Formal experiment name in use:

- `qwen3_8b_grpo_vllm_azurejudge_full`

### D. Base and RL inference runs already completed

Confirmed scratch inference outputs:

- Task 1 base:
  - `/gpfs/radev/scratch/cohan/yz979/xucai/Abforge_Training/infer/task1_base_qwen3_infer_task1.jsonl`
- Task 1 API checkpoint:
  - `/gpfs/radev/scratch/cohan/yz979/xucai/Abforge_Training/infer/task1_api_ckpt14_qwen3_infer_task1.jsonl`
- Task 2 API checkpoint:
  - `/gpfs/radev/scratch/cohan/yz979/xucai/Abforge_Training/infer/task2_api_ckpt14_qwen3_infer_task2.jsonl`
- historical Task 2 base:
  - `/gpfs/radev/scratch/cohan/yz979/xucai/Abforge_Training/infer/base_qwen3_infer_task2.jsonl`
- historical Task 2 RL checkpoint:
  - `/gpfs/radev/scratch/cohan/yz979/xucai/Abforge_Training/infer/ckpt28_qwen3_infer_task2.jsonl`

Merged HF checkpoint directories already present:

- `/gpfs/radev/scratch/cohan/yz979/xucai/Abforge_Training/infer/task1_api_ckpt14_merged_hf`
- `/gpfs/radev/scratch/cohan/yz979/xucai/Abforge_Training/infer/task2_api_ckpt14_merged_hf`

### E. Local judge historical run

Status:

- local judge worked functionally but was very slow and GPU-inefficient

Known historical local Task 2 checkpoint root:

- `/gpfs/radev/project/cohan/yz979/xucai/Abforge_Training/verl_proj/checkpoints/abforge_task2/qwen3_8b_grpo_vllm`

Known merged / inferred artifact from that line:

- `/gpfs/radev/scratch/cohan/yz979/xucai/Abforge_Training/infer/ckpt28_qwen3_infer_task2.jsonl`

Recommendation:

- do not choose local judge for new formal runs unless the user explicitly asks for it

## 22. Quick Commands

These commands are safe defaults to suggest when the user asks "what do I run next?"

### Set up shell variables

```bash
export CODE_ROOT=/gpfs/radev/project/cohan/yz979/xucai/Abforge_Training
export WORK_ROOT=/gpfs/radev/scratch/cohan/yz979/xucai/Abforge_Training
export BASE_MODEL=$WORK_ROOT/models/Qwen3-8B
```

### Activate main env

```bash
module reset
module load miniconda
module load GCCcore/13.3.0
module load CUDA/12.4.1
source $(conda info --base)/etc/profile.d/conda.sh
conda activate $WORK_ROOT/conda/abforge_vllm
```

### Start Task 2 API reward

```bash
cd $CODE_ROOT
sbatch slurm/task2_azure_reward.slurm
```

### Start Task 1 API reward

```bash
cd $CODE_ROOT
sbatch slurm/task1_azure_reward.slurm
```

### Watch a job log

```bash
tail -f $WORK_ROOT/logs/<log_name>.err
```

### Check a job state

```bash
squeue -j <JOBID>
sacct -j <JOBID> --format=JobID,JobName,State,ExitCode,Elapsed,NodeList
```

### Merge an FSDP actor checkpoint

```bash
python3 $CODE_ROOT/verl_proj/scripts/model_merger.py \
  --backend fsdp \
  --hf_model_path $BASE_MODEL \
  --local_dir <global_step_dir>/actor \
  --target_dir <merged_hf_dir>
```

### Run local inference

```bash
python3 $WORK_ROOT/run_inference_local.py \
  --task <1_or_2> \
  --input $WORK_ROOT/data/bench_data_4_subset50_simple.jsonl \
  --output-prefix <prefix> \
  --model-path <hf_model_dir> \
  --tokenizer-path $BASE_MODEL \
  --trust-remote-code \
  --dtype bf16 \
  --device-map auto \
  --overwrite
```

## 23. Decision Guide For Future Codex Sessions

This section is the practical default policy layer.

When a future Codex session starts, use these rules unless the user explicitly overrides them.

### A. If the user says "formal training"

Default interpretation:

- use API judge, not local judge
- use full parquet, not small test parquet
- use base Qwen3-8B actor as the training start point
- use scratch checkpoint directory

Task mapping:

- Task 1 formal:
  - `$WORK_ROOT/data/abforge_task1_rl_full/train.parquet`
  - `$WORK_ROOT/data/abforge_task1_rl_full/val.parquet`
  - `slurm/task1_train_vllm_azurejudge.slurm`
- Task 2 formal:
  - `$WORK_ROOT/data/abforge_task2_rl_full/train.parquet`
  - `$WORK_ROOT/data/abforge_task2_rl_full/val.parquet`
  - `slurm/task2_train_vllm_azurejudge.slurm`

Formal defaults:

- `trainer.total_epochs=1`
- `trainer.save_freq=100`
- checkpoint path under `$WORK_ROOT/checkpoints/..._full`
- experiment name suffixed with `_full`

### B. If the user says "test training"

Default interpretation:

- use API judge
- use small parquet
- do not overwrite formal experiment names

Task mapping:

- Task 1 test:
  - `$WORK_ROOT/data/task1/train.parquet`
  - `$WORK_ROOT/data/task1/val.parquet`
- Task 2 test:
  - `$WORK_ROOT/data/task2/train.parquet`
  - `$WORK_ROOT/data/task2/val.parquet`

### C. If the user says "resume training"

Default interpretation:

- keep the same checkpoint root
- keep the same experiment name
- inspect whether the directory already has `latest_checkpointed_iteration.txt`
- use `resume_mode=auto` behavior rather than inventing a new checkpoint path

Resume checklist:

1. identify whether the checkpoint is under `project` or `scratch`
2. identify the exact experiment directory
3. inspect:
   - `latest_checkpointed_iteration.txt`
   - available `global_step_xxx` directories
4. only then adjust `trainer.total_epochs`

For "continue from epoch 1 to epoch 2", the default action is:

- keep `trainer.default_local_dir` unchanged
- change `trainer.total_epochs=1` to `trainer.total_epochs=2`

### D. If the user says "run inference"

Default interpretation:

- use the scratch inference script:
  - `$WORK_ROOT/run_inference_local.py`
- use the scratch benchmark file:
  - `$WORK_ROOT/data/bench_data_4_subset50_simple.jsonl`

Then decide:

- base model inference:
  - use `$WORK_ROOT/models/Qwen3-8B`
- RL checkpoint inference:
  - first merge `global_step_xxx/actor`
  - then point `--model-path` at the merged HF directory

Never assume:

- `run_inference_local.py` is in Git on Misha
- checkpoint is already merged

### E. If the user says "use local judge"

Default interpretation:

- pause and confirm they really want it

Reason:

- local judge historically worked but was much slower
- API judge is the current default for formal work

If the user still wants local judge:

- inspect whether they mean Task 1 or Task 2
- prefer the existing Misha slurm files under `$CODE_ROOT/slurm`
- verify model path and resource sizing before launch

### F. If the user says "compare base vs RL"

Default interpretation:

- infer all requested models on the same benchmark input
- write outputs under `$WORK_ROOT/infer`
- use clear prefixes like:
  - `task1_base_qwen3`
  - `task1_api_ckpt14_qwen3`
  - `task2_base_qwen3`
  - `task2_api_ckpt14_qwen3`

### G. If something looks missing

Default troubleshooting order:

1. check whether the file exists under `project`
2. check whether it actually lives under `scratch`
3. check whether it is server-only and untracked
4. only then conclude it is missing

Examples:

- `run_inference_local.py` exists on scratch
- slurm files exist under project root
- secrets exist under scratch

### H. If a path assumption is uncertain

Default behavior:

- do not guess silently
- ask the user for a short server command if needed
- prefer commands like:
  - `find ...`
  - `ls ...`
  - `grep ...`

This repo has enough server-only state that path verification is often necessary.

## 24. Do Not Store

Do not write these into Git:

- Azure API keys
- server-only secrets
- user-specific scratch-only temporary artifacts

Keep secrets only in:

- `$WORK_ROOT/secrets/*.env`