This page compiles the run. It does not run it.

Memory floor

The accounted components fit, but 5 components could not be computed from these inputs. Whether the run fits is undetermined, and no arrangement of unknowns makes it a yes.

5 components are undetermined, so the accounted floor is lower than the real requirement by an unknown amount. Fill in the architecture dimensions above to close the gap.

Guardrails

• MaterialThe Hub dataset is not pinned to a revision

  A Hub dataset repository is mutable in exactly the way a model repository is. Pin it with `--revision <sha>` and record that SHA, or download it once and record the SHA-256 of the bytes on disk. Otherwise 'trained on this dataset' names a moving target.

  Source ↗

• MaterialNo dataset content digest

  The receipt binds the run to the exact bytes the training loop read. Compute `shasum -a 256` over the file the run will open and carry that digest through to the receipt. A dataset name is not a dataset.

• MaterialOverlap with the benchmarks you intend to report is unchecked

  If any part of this dataset overlaps a benchmark you later quote, the number you quote measures memorisation rather than capability. Instruction sets assembled from the open web routinely contain benchmark items verbatim, and synthetic data generated by a model that saw those benchmarks reproduces them at a rate nobody's upload flow measures. Run an n-gram overlap pass against every evaluation you plan to report, and record the result — including a null result — in the receipt.

• MaterialNo held-out split

  Without a split the training loop never sees, a falling loss curve is a statement about fit to the training set and nothing else. Hold data out before the first run, not after a result you like.

• MaterialA run without a baseline cannot tell you whether it helped

  Evaluate the pinned base checkpoint on the same harness, at the same version, with the same prompts and decoding parameters, before training starts — and keep the raw output. A post-training score with no matched pre-training score on the same harness is uninterpretable: it cannot distinguish a real improvement from a prompt-format change, a harness version bump or noise. Report both numbers with their error bars or report neither.

• MaterialThe output inherits every licence in the chain

  A fine-tuned model is a derivative work of everything that went into it. This run's output is constrained by the base model's licence (apache-2.0), the dataset's licence. The most restrictive term in that chain governs the result — not the most convenient one — and a permissive base does not launder a restricted dataset.

• NoteHalf the public fine-tuning corpus points at stopped projects

  Axolotl was verified live on 2026-09-05 against its own repository. torchtune wound down in 2025 and its README says so; Hugging Face AutoTrain Advanced states that it is no longer maintained, that no new features will be added and that bugs will not be fixed. Both are still the target of tutorials, course material and model cards published after they stopped. Check the liveness matrix below before following any fine-tuning guide.

  Source ↗

Compiled recipe

# MOEModels training recipe
# Base repository : Qwen/Qwen3-0.6B
# Base revision   : c1899de289a04d12100db370d81485cdf75e47ca
# Method          : QLoRA (4-bit base)
# Framework       : Axolotl 0.18.0
# Dataset         : HuggingFaceH4/ultrachat_200k (train)
# Dataset digest  : NOT COMPUTED
# Seed            : 3407
#
# This file carries no timestamp and no random value, so identical inputs produce identical bytes. That is what makes the digest in recipe.lock.json worth recording.
# MOEModels compiled this run. It did not run it, and it makes no claim about the result.

base_model: "./base/Qwen3-0.6B"
sequence_len: 4096

adapter: qlora
lora_r: 32
lora_alpha: 64
lora_dropout: 0.0
lora_target_linear: true

datasets:
  - path: "HuggingFaceH4/ultrachat_200k"
    split: "train"
    type: "chat_template"
val_set_size: 0.05

output_dir: "./out/Qwen3-0.6B-qlora"
micro_batch_size: 1
gradient_accumulation_steps: 8
num_epochs: 2
learning_rate: 0.0002
lr_scheduler: cosine
warmup_ratio: 0.03
optimizer: adamw_bnb_8bit
gradient_checkpointing: true
bf16: true
fp16: false
seed: 3407
logging_steps: 10
saves_per_epoch: 1
evals_per_epoch: 2
load_in_4bit: true

# Versions MOEModels did not read from a package index are left unpinned rather than
# guessed. After the first successful install, freeze the environment and record it.

Commands

# 1. Authenticate. Required for a gated repository, harmless otherwise.
hf auth login --token "$HF_TOKEN"
# 2. Pin the base checkpoint to an immutable revision.
hf download Qwen/Qwen3-0.6B --revision c1899de289a04d12100db370d81485cdf75e47ca --local-dir base/Qwen3-0.6B
# 3. Pin the dataset and record the bytes the run will actually read.
hf download HuggingFaceH4/ultrachat_200k --repo-type dataset --local-dir data/ultrachat_200k   # unpinned: add --revision <SHA>
find data/ultrachat_200k -type f -exec shasum -a 256 {} \; | tee dataset.sha256
# 4. Pin the toolchain. Unpinned entries are ones MOEModels did not verify.
python -m venv .venv && . .venv/bin/activate
pip install "axolotl==0.18.0" datasets accelerate bitsandbytes
pip freeze > environment.lock.txt
# 5. Measure the base BEFORE training, on the harness you intend to report.
#    Keep the raw output. A post-training score without this is uninterpretable.
# 6. Record the recipe digest that the receipt will carry.
shasum -a 256 recipe.lock.json
# 7. Run, with an explicit wall-clock ceiling.
timeout 10800 axolotl train axolotl.yaml 2>&1 | tee train.log
# 8. Measure the result on the same harness, at the same version, same prompts.
# 9. Publish, or keep it local.
# result saved to ./out/Qwen3-0.6B-qlora — no Hub repository configured

Recipe fingerprint

0726cf1ca8deec5eNon-cryptographic. It detects accidental drift between two recipes; it is not a security property and must not be presented as one.

Recipe digest (SHA-256)

Unavailable in this context. WebCrypto is absent, so compute it where the recipe runs with shasum -a 256 recipe.lock.json.