使い方
- インストールコマンドをコピーし、自分の環境へ導入する前にソースを確認します。
- 必要なツール、認証情報、対応環境を原文で確認します。
- 入力と成果物を指定し、出力を資料と照合します。
npx skills add https://github.com/modelstudioai/cli --skill bailian-opd開始前の確認
- 導入にはモデル契約、API枠、ファイルへのアクセス許可は含まれません。
- スター数や導入数は活動の指標であり、独立した品質・安全性試験ではありません。
見出し、例、コード、表、リンク、参照画像を含む原文を表示しています。
Bailian OPD Distillation Workflow
Create a pure Model OPD training job (training_type=pg_opd, top-level teacher_model, no rollout/reward functions, no Function Compute) with the highest possible first-run success rate. Gate logic lives in deterministic scripts under scripts/; the agent orchestrates, presents findings, and obtains consent — it does not re-implement gates or guess flags.
CRITICAL — shared protocol first. Before executing anything, Read../bailian-protocol/SKILL.mdand follow it: mandatory version pre-flight before the firstblcall, consent classes, attribution, error reporting, output conventions. If the protocol skill is missing, stop and promptbl skill init. Business context for OPD flags lives in skillbailian-finetune(itsreference/finetune.md) orbl finetune opd --help— never guess flags.
Attribution
Bundled scripts already carry the protocol script attribution (BAILIAN_CALLER_TYPE=script, BAILIAN_CALLER_NAME from --caller-name). For inline bl commands the agent runs itself, use the protocol agent template, e.g.:
BAILIAN_CALLER_TYPE=agent BAILIAN_CALLER_NAME=codex BAILIAN_SKILL_NAME=bailian-opd BAILIAN_SKILL_VERSION=2.10.2 bl finetune opd authorize --authorize --yesPrerequisites
blinstalled and up to date (protocol version pre-flight);node>= 18 on PATH.- Two logins: API key (
bl auth login --api-key) for data/params/create, and
Console (bl auth login --console) for the OPD feature authorize probe.
- OPD is invite-only (邀测): the account whitelist must be enabled. Published
student/teacher pairs and sizing: the official OPD docs "模型与训练单元" section (the CLI warns on unpublished pairs).
Workflow (steps 0–7)
0. Protocol pre-flight
Run the protocol version pre-flight; check node --version (>= 18) and bl auth status. Missing Console login blocks only the authorize stage — the preflight report still surfaces everything else.
1. Chain-start authorization
Disclose before any probe: OPD training is MTU postpaid billing (mtu_spec_code × mtu_capacity, charged while the job runs). Obtain the user's consent to start the read-only preflight. This consent covers the authorize probe's --yes (the runtime risk gate requires it even for the read-only probe; the probe never enables anything).
2. Choose a baseline (do not improvise hyper-parameters)
Default to the all-explicit 9B←397B baseline for a first run — every teacher resource field is explicit, so nothing depends on server defaults and capacity is fully computable:
node scripts/opd-preflight.mjs \
--base-model qwen3.5-9b --teacher-model qwen3.5-397b-a17b \
--datasets <train.jsonl> --validations <valid.jsonl> \
--hyper-parameters-file assets/baseline-9b-from-397b.hyper_parameters.json \
--resource-config-file assets/baseline-9b-from-397b.resource_config.json \
--caller-name <host>Second choice: assets/baseline-9b-from-flash.* (9B←qwen3.8-flash, teacher fields omitted → server fills Flash defaults). Capacity guide:
| Layout | mtu_capacity |
|---|---|
| Default student layout (layout fields omitted), 9B + Flash teacher | 24 (official recommendation; observed default = 16 student + 8 teacher GPUs) |
| Flash baseline custom layout (4/2 GPUs) | 16 |
| 397B baseline all-explicit (8/8 GPUs + 8 teacher) | 24 |
Baselines are dated snapshots of historically successful requests — never "improve" them; apply overrides only when the user explicitly asks. Data format: JSONL, messages with system (task) + user (input); prompt-only is valid for pure distillation; rollout_extra.solution passes through untouched. Data size: batch_size is the samples-per-step, so the training set needs at least ceil(batch_size / n_epochs) effective samples for one step — the preflight batch-fit stage computes the step count and blocks zero-step configurations (they fail at execution time, after billing starts).
3. Preflight report and remediation
Run opd-preflight.mjs (read-only; aggregates all stages in one pass). Present the report. Blocked stages → follow their remediation:
- Console login missing →
bl auth login --console. - Service roles
REQUIRES_USER_ACTION→ surface the manual authorization links
verbatim; the CLI never creates roles.
business-enabled: MISSING→ only after explicit user confirmation, run
inline: bl finetune opd authorize --authorize --yes. The scripts never enable.
- Empty parameter schema → the student is not registered for pg_opd; pick a
published student.
- Workspace 400/403 → the workspace does not match the API key; fix
--workspace-id / config.
batch-fitblocked (0 training steps) → add data to at least
ceil(batch_size / n_epochs) samples with headroom, or lower batch_size / raise n_epochs; re-run preflight.
Re-run preflight after remediation until no stage is blocked.
4. Consent gate (mandatory, verbatim)
Present to the user, verbatim from the report:
body_preview(the exact request body),- every warning (
OPD_PAIR_NOT_PUBLISHED,OPD_DEFAULT_LAYOUT_CAPACITY,
OPD_LENGTH_EXCEEDS_TEACHER, OPD_BATCH_FIT_UNKNOWN, OPD_WARMUP_EXCEEDS_STEPS, OPD_INTERVAL_EXCEEDS_STEPS, capacity advice, dataset warnings),
pending_uploads(local files that will be uploaded),- billing scope:
mtu_spec_code × mtu_capacity, MTU postpaid, charged until the
job reaches a terminal state.
Obtain explicit confirmation. Any scope change (model, teacher, data, capacity, hyper-parameters) requires re-running preflight and re-confirming. Never add --yes or --confirmed automatically.
5. Confirmed submit
node scripts/opd-submit.mjs <same assembly flags> --confirmed --caller-name <host>The script passes --yes to bl finetune opd run only because --confirmed was given after step 4. Local paths are validated and uploaded inside this run. On create failure/timeout the script never retries; it runs a duplicate-creation guard (the server can persist a job and still return 500) and reports duplicate_check. possibly_created → attach to that job, do not resubmit. unknown → stop and report the uncertainty.
6. Monitor to a verified terminal state
node scripts/opd-wait.mjs --job-id <ft-xxx> [--interval 30] [--poll-timeout 3600] --caller-name <host>(or opd-submit.mjs … --confirmed --wait). Creation returning a job id is NOT success: capacity mismatches fail at execution time, after billing starts. Poll timeout (exit 5) ≠ failure — the job was not canceled; resume with the printed command, never resubmit.
7. Verified success and hand-off
Success = terminal SUCCEEDED and checkpoints listed (the wait script includes them). Give the protocol one-line summary. Export / deployment / metrics → soft hand-off to skill bailian-finetune (bl finetune checkpoints / export / metrics, then bl deploy). Listing jobs: bl finetune list (there is no opd list).
Failure triage
| Signal | Meaning | Action |
|---|---|---|
| exit 2 | usage / local guard (missing MTU fields, bad charge_type, TP ≠ GPUs/node, 0 training steps, mini-batch too large, schema violation) | fix inputs; nothing was uploaded or created |
| exit 3 | auth missing | bl auth login --api-key / --console |
| exit 5 | poll timeout | job still alive; resume waiting |
| exit 6 | network | retry the read-only step; never blind-retry create |
| exit 7 | requires_confirmation | normal control flow — obtain consent, then --yes/--confirmed |
| create 500 / timeout | job may exist anyway | trust duplicate_check; attach or stop-and-report |
| FAILED message "OPD requires N student GPUs + M teacher GPUs, but the MTU capacity is K" | execution-phase capacity mismatch | set capacity N+M (or change layout), re-preflight, re-confirm, submit as a NEW job |
| batch-fit blocked / exit 2 "0 training steps" | samples × nepochs < batchsize (samples per step) | add data to ≥ ceil(batchsize / nepochs) with headroom, or lower batchsize / raise nepochs; re-preflight, re-confirm |
| FAILED length errors | student lengths exceed teacher limit | keep max_length/max_prompt_length ≤ teacher max_model_len (Flash/Max default 16385); re-confirm |
Anti-patterns (hard rules)
- Never fill the parameter API's
recommend_valueinto requests — it is not the
selected teacher's effective default (Flash/Max teacher max length 16385 vs the generic 33792/34816).
- Never hardcode teacher defaults into requests; omitted teacher fields are
resolved by the server.
- Never pass
--yestoopd runor--confirmedto the submit script without
explicit user confirmation of the exact scope.
- Never blind-retry a create; a timeout is not a rejection.
opd authorizeexit 0 does NOT mean enabled — parseok/checks.- Treat capacity advice as advisory: teacher mode (internal/external) has no
public query API, so the scripts compute both interpretations and warn.
References
- Shared execution protocol:
../bailian-protocol/SKILL.md - OPD command flags and domain rules: skill
bailian-finetune,bl finetune opd --help - User-facing intro:
README.md/README.zh.md