SlowGuess/abforge-isambard-transfer

Status (2026-07-30): retired transfer bundle — do not use as a model source. The Isambard migration did not proceed; the unified run finished on the original cluster. This bundle no longer reflects a usable training state:

• Removed: the abandoned 0.66-epoch qwen3_8b_sft_combined/global_step_400 checkpoint (the SFT it seeded produced a degenerate RL run) and the paper draft PDF.
• Superseded: the mixed SFT-602 checkpoint is now published as a proper model repo, SlowGuess/ABForge-Qwen3-8B-Combined-SFT. Use that instead of the raw shards here.
• The released unified model is SlowGuess/ABForge-Qwen3-8B-Combined (RL step 200).

The sections below are the original transfer notes, kept for the record. Their instructions reference checkpoints that have been deleted.

ABForge Isambard transfer

Private transfer bundle for continuing combined SFT+RL training on Isambard.

Contents:

• code/Abforge_Training/: training code snapshot from Misha project checkout, excluding git/cache/runtime artifacts.
• data/abforge_combined_rl/: combined RL train/val parquet.
• checkpoints/abforge_combined_sft/qwen3_8b_sft_combined/global_step_400/: SFT400 init checkpoint.
• checkpoints/abforge_combined/qwen3_8b_grpo_combined_from_sft400/global_step_40/: RL resume checkpoint.
• checkpoints/abforge_combined/qwen3_8b_grpo_combined_from_sft400/latest_checkpointed_iteration.txt: resume pointer set to 40.

Not included:

• DeepSeek-V4-Flash HF cache. Download separately on target cluster.
• Secrets such as W&B or HF tokens.
• Misha/Bouchet conda environments; target cluster must build its own ARM/GH200 environment.

t1x2 update (task1 up-weighted RL) — added 2026-07-07

Combined task1 lags the single-task model (bench C-score ~54.5 peak @ SFT->RL step180 vs single-task 58.7). Root cause = a task1 RECALL/GT-alignment gap that RL instills, and combined RL dilutes it (half the gradient is task2). Fix under test: oversample task1 2x in the RL batch (task1:task2 = 1.9:1 on train; val stays 1:1 for unbiased eval).

New in this bundle:

• data/abforge_combined_rl_t1x2/: the up-weighted RL data (train task1=56894 / task2=29900; val 100/100 unchanged). This is abforge_combined_rl with task1 rows duplicated once.
• Refreshed code/.../abforge_combined_merge.py (adds --task1_repeat) and code/.../slurm/_inner_combined_rl.sh (honors COMBINED_RL_DIR to swap the RL dataset).

Run on Isambard (start from the SFT-400 init already in this bundle):

# point the combined-RL launcher at the t1x2 data + SFT-400 init
INIT_MODEL=<ISAMBARD>/checkpoints/abforge_combined_sft/qwen3_8b_sft_combined/global_step_400
COMBINED_RL_DIR=<ISAMBARD>/data/abforge_combined_rl_t1x2
EXPERIMENT_NAME=qwen3_8b_grpo_combined_from_sft_t1x2
TOTAL_STEPS=200  SAVE_FREQ=20  TEST_FREQ=20  MAX_CKPT_KEEP=4
# _inner_combined_rl.sh reads COMBINED_RL_DIR; the data is pre-built so it skips the merge.

Watch: val task1 recall should climb toward 0.54 (single-task level) vs the balanced run's ~0.47. Trade-off: may cost a little task2 (has ~0.02-0.03 headroom vs single-task 0.614). The same job is also queued on Misha as qwen3_8b_grpo_combined_from_sft_t1x2.

SFT-602 + API-judge RL (2026-07-15) — the corrected combined pipeline

Root cause of combined task1 lag was traced to (a) the combined SFT ran only 0.66 epoch (TIMEOUT) and (b) the RL used the LOCAL open-weights judge (which rewards template/umbrella hacks the DeepSeek API judge suppresses). Fix = complete the SFT + switch RL to the API judge.

New in this bundle:

• checkpoints/abforge_combined_sft/qwen3_8b_sft_combined_full1ep/global_step_602/: the COMPLETED full-epoch combined SFT. bench_50: task1 C=35.1 (> single-task SFT 31.9), task2 design 0.512. Use this as the RL init (NOT the old 0.66-epoch SFT-400).
• code/.../slurm/combined_rl_apijudge.slurm: reference launcher (adapt modules/paths for GH200).

Run on Isambard (SFT-602 + DeepSeek API judge, matches single-task's recipe except task2 is still in the batch):

# 1) put the DeepSeek key in secrets/judge_api.env:
#    JUDGE_API_BASE=https://api.deepseek.com   JUDGE_MODEL=deepseek-v4-flash   JUDGE_API_KEY=sk-...
# 2) the combined reward server (reward_part/combined_judge/) calls the API directly -> NO local
#    judge node needed, just the training GPUs.
INIT_MODEL=<ISAMBARD>/checkpoints/abforge_combined_sft/qwen3_8b_sft_combined_full1ep/global_step_602
COMBINED_RL_DIR=<ISAMBARD>/data/abforge_combined_rl     # balanced (task1:task2 = 1:1)
EXPERIMENT_NAME=qwen3_8b_grpo_combined_sftfull_apijudge
LR=1.5e-5   TOTAL_STEPS=200   SAVE_FREQ=20   TEST_FREQ=20   MAX_CKPT_KEEP=6
# judge=API deepseek-v4-flash (SECRETS_FILE=secrets/judge_api.env), thinking disabled, temp=0

Watch: val task1 -> single-task's API-val level (0.58) AND task2 val holding (0.61). The only irreducible diff vs single-task task1 sft+RL is that task1 is 50% of the batch (multi-task dilution) — hence 200 steps (~2x single's 100) for task1-exposure parity. Mind DeepSeek balance (each 200-step run costs a few hundred CNY; API 402 mid-run silently zeros the reward). The same run is also live on Misha (job 2116006).