laion/axolotl-ota-sft-integration

Axolotl ⇄ OpenThoughts-Agent SFT-backend integration — overview

Status: COMPLETE + merged to penfever/working (merge 02d676d0, 2026-07-02). Where it ran: TACC Vista (GH200, aarch64), conda env sft-axolotl. One-line result: OpenThoughts-Agent can now run SFT through axolotl (--sft_backend axolotl) as a drop-in alternative to LLaMA-Factory, with the delphi chat-template masking validated on the real delphi path (jinja-as-ground-truth: train == serve).

---

1. What was done

Two staged design→execute cycles (plans + per-stage scopes under notes/):

Cycle 1 — marin axolotl fork with 3 ported features (notes/cycle1_marin-fork-3feature-port/). Created marin-community/axolotl @ feuer/marin-fork-3feature-port (3c206072, axolotl 0.17.0.dev0):

• delphi.jinja chat template (auto-globbed) — the Llama-3 turn format + reasoning/tool tokens (<|start_think|>/<|end_think|>/<|tool_call|>/<|tool_result|>).
• template_integrity plugin — the save-time footgun fix: keeps a populated chat_template embedded in tokenizer_config.json on every checkpoint save (tokenizer_save_jinja_files: false + per-checkpoint coverage) so the model is not silently OOD at serve time (the historical 0%-SWE-bench bug).
• mfu plugin (MFU logging) + supabase_registry plugin (opt-in train-end model registration).

Cycle 2 — wire axolotl into hpc.launch as an SFT backend (notes/cycle2_sft-launch-backend/). --sft_backend {llamafactory,axolotl} (LF default, flag-off byte-identical). Submodule sft/axolotl pinned at 3c206072. Runner dispatch (-m axolotl.cli.train), an LF-exp-args → axolotl-YAML translator (hpc/axolotl_config_utils.py), config assets (sft/axolotl_configs/), and validation gates (sft/axolotl_gates/). Validated on TACC across 6 stages.

---

2. Gate results (see results/ for the raw artifacts)

Gate	Verdict	Evidence
Stage 3 — end-to-end smoke	✅ GO	job 801458; loss 2.272→0.951 (6 steps); SDPA; ckpt w/ embedded delphi template
Stage 4 — footgun-through-launcher	✅ GO	embedded chat_template byte-identical to canonical delphi.jinja across output + checkpoint dirs
Stage 6 — LF-vs-axolotl loss-match	❌ NO-GO (understood)	axolotl 1.567→1.549, LF 1.334→0.702 (both 30 steps); 120% gap = a framework turn-masking divergence on the non-canonical guanaco/llama3 dataset, not a defect
Delphi masking canary (the check that matters)	✅ PASS	job 802053; loss 2.628→2.532 (20 steps); <|start_think|>…<|end_think|>+assistant trained, user/system masked, 0 "Last turn is not trainable" skips; trainable fraction 73–96% (via axolotl.cli.preprocess --debug)

Takeaway: the Stage-6 guanaco gap was a masking divergence specific to that non-canonical dataset. On the real delphi path the axolotl delphi template masks correctly — jinja-as-ground-truth holds.

Delphi mask-dump evidence (job 802053, axolotl.cli.preprocess --debug)

Dump format text(label, token_id): label -100 = masked, label == token_id = trained. Consistent across 4/4 examples:

Assertion	Result
<|start_think|>/<|end_think|> reasoning span trained	(128002,128002) + (128003,128003) TRAINED
assistant answer trained	The(791,791) answer(4320,4320) is(374,374)… TRAINED
user turns masked	user(-100,882) + content (-100,…) MASKED
system turns masked	<|begin_of_text|>, system header+content (-100,…) MASKED
assistant header masked (only content trained)	assistant(-100,78191) MASKED
Llama-3 structure segmented	<|start_header_id|>…<|eot_id|> boundaries correct
Last turn is not trainable, skipping	0 occurrences

Trainable-token fractions per example: 659/763 (86%), 1420/1474 (96%), 636/726 (88%), 463/631 (73%) — high, as expected for reasoning data (short masked prompt, long trained reasoning+answer). One benign nuance: the trailing <|eot_id|> of the last assistant turn is masked (a train_on_eos policy choice, not a defect).

Loss series (from results/trainer_states/):

• axolotl_parity_llama3_ckpt30.json — 30 steps, 1.5667→1.5490
• lf_parity_llama3_ckpt30.json — 30 steps, 1.3342→0.7020
• delphi_canary_ckpt20.json — 20 steps, 2.6280→2.5315
• axolotl_smoke_ckpt6.json — 6 steps, 2.2722→0.9505

---

3. Fixes shipped along the way

Axolotl backend / launcher (7): submodule + --sft_backend selector; runner dispatch; LF→axolotl config translator; CLI global_batch_size int-cast; short job-scoped TMPDIR (AF_UNIX 108-byte sun_path fix, length-conditional, BOTH backends); expandable_segments allocator (BOTH backends); the torchao==0.17.0 aarch64 env fix (doc, not a submodule patch).

LLaMA-Factory transformers-5.x (4, bonus — unbreak LF SFT on modern transformers): LF-fork d20b8666 = add_special_tokens(replace_additional_special_tokens=…) signature-guard; launcher: report_to=none on no-internet nodes (wandb-0.28 service-socket crash), the TMPDIR + expandable_segments generalizations above. LF pin bumped 6617a420 → d20b8666.

---

4. Example launch commands

4a. Axolotl SFT — the validated delphi masking canary (TACC Vista, aarch64)

# prereq once: prep the Delphi tokenizer (single delphi tokens)
python sft/delphi/prepare_delphi_tokenizer.py \
  --model laion/delphi-3e18-p33m67-k0p20-lr83-a003 --output $SCRATCH/delphi-canary-tok

# launch (WANDB_MODE=disabled: TACC is an internet node -> report_to=wandb; wandb 0.28 crashes offline)
WANDB_MODE=disabled python -m hpc.launch --job_type sft --sft_backend axolotl \
  --train_config_path sft/axolotl_configs/delphi_canary.yaml \
  --model_path $SCRATCH/delphi-canary-tok --conda_env sft-axolotl \
  --dataset open-athena/llama-nemotron-science-reasoning-on-le3000tok-100k-canonical-think \
  --messages messages --role_tag role --content_tag content \
  --partition gh-dev --num_nodes 1 --gpus_per_node 1 --time_limit 01:00:00

Launcher gotcha: hpc.launch REBUILDS the axolotl datasets: block from --dataset + --messages/--role_tag/--content_tag; the in-config datasets: block is honored ONLY by direct axolotl.cli.preprocess. So pass those dataset flags on the launcher path.

4b. Mask-dump validation (the decisive masking check — no training)

python -m axolotl.cli.preprocess sft/axolotl_configs/delphi_canary.yaml \
  --base_model $SCRATCH/delphi-canary-tok --debug   # via srun (login node OOMs on tokenizer loads)
# inspect: <|start_think|>..<|end_think|> + assistant TRAINED (label==id), user/system MASKED (-100),
#          0 "Last turn is not trainable, skipping"

4c. Axolotl vs LLaMA-Factory SFT (backend swap — same launcher, same flags)

# axolotl backend
python -m hpc.launch --job_type sft --sft_backend axolotl  --conda_env sft-axolotl \
  --train_config_path sft/axolotl_configs/<cfg>.yaml --dataset <ds> --messages messages \
  --role_tag role --content_tag content --num_nodes 1 --gpus_per_node 1 --time_limit 02:00:00
# llamafactory backend (default; omit --sft_backend)
python -m hpc.launch --job_type sft --sft_backend llamafactory --conda_env otagent \
  --train_config_path sft/lf_configs/<family>/<cfg>.yaml --dataset <ds> \
  --role_tag role --user_tag user --assistant_tag assistant --content_tag content \
  --num_nodes 1 --gpus_per_node 1 --time_limit 02:00:00

4d. Downstream — RL the SFT'd delphi model on CoreWeave / Iris (the next pipeline stage)

The axolotl SFT backend produces the instruction-tuned delphi checkpoint that then goes to GRPO RL on the CoreWeave H100 GPU cluster (via the marin Iris SDK). Representative (see skill rl-agentic-launch-iris):

python -m rl.cloud.launch_rl_iris \
  --rl_config <rl_yaml> --model_path laion/<sft-delphi-ckpt> \
  --train_data <task_parquet> --num-nodes 1 \
  --rendezvous-dir <gs://…/rendezvous> --job-name <name> --priority normal --cpu 48 --max-retries 1

This SFT→RL handoff is why the delphi masking correctness matters: train==serve on delphi.jinja means the SFT'd model isn't OOD when it hits the RL rollouts + eval.

---

5. Reproduce / investigate — directory guide

• OVERVIEW.md — this file.
• agent_logs/ — the full dated execute logs for both cycles (blow-by-blow debug record).
• notes/cycle1_*, notes/cycle2_* — the staged design plans + per-stage scopes + the DEFERRED-TACC checklist.
• configs/ — the axolotl configs (delphi_canary.yaml, smoke.yaml, axolotl_parity_llama3.yaml, marin_delphi_all3.yaml), the LF parity config, and the two validation gate scripts (stage4_footgun_through_launcher.py, stage6_parity.py).
• results/trainer_states/ — the loss series (JSON) for smoke / both parity sides / the delphi canary.
• results/rendered_configs/ — the exact launcher-rendered axolotl train config for the canary (byte repro).
• results/run_logs/ — the training .out logs (env fixes engaged, SDPA, template embed, loss lines) for the axolotl parity (801488), LF parity (801998), and the delphi canary (802053).

6. Key references (in-repo)

• Backend wiring: hpc/sft_launch_utils.py, hpc/axolotl_config_utils.py, hpc/arguments.py.
• Configs/gates: sft/axolotl_configs/, sft/axolotl_gates/, sft/delphi/prepare_delphi_tokenizer.py, sft/delphi/dataset_info.json.
• Dependency facts + gotchas: .claude/projects/axolotl/axolotl.md.
• Skill: .claude/skills/sft-launch/ (merged jupiter+leonardo; backend + delphi guidance).
• Fork: marin-community/axolotl @ feuer/marin-fork-3feature-port (3c206072).