File size: 25,620 Bytes
c2295ac
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
77456ce
c2295ac
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
77456ce
c2295ac
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
# MVP build for “Data Curation Workbench” (Hugging Face Space)

## 0) MVP Goal & Scope

**Goal:** Let a signed‑in user upload **D₀** (or reference a Hub dataset), pick a **model** + **metrics**, choose candidate datasets **{D₁…Dₙ}**, launch **small‑scale fine‑tunes/evals** as detached **Jobs**, and view:

* per‑run metrics (loss / F1 / Exact‑Match),
* a **scaling‑law** plot, and
* a table ranking which Dₖ helps the most,
* with all artifacts saved to a results dataset or Space storage.

**Out of scope (for MVP):**

* Multi‑GPU distributed training, multi‑task mixing UI, complex hyperparam sweeps.
* Non‑text tasks.

---

## 1) Repository Layout

Create these files/folders:

```
.
├─ README.md
├─ PLAN.md                        # this file
├─ app.py                         # Gradio UI + Job submission + status polling
├─ requirements.txt
├─ catalog/
│  └─ candidates.json             # curated {D₁…Dₙ}
├─ utils/
│  ├─ hub.py                      # upload to Hub, results repo helpers
│  ├─ data.py                     # dataset loading/mixing/helpers
│  └─ plotting.py                 # scaling plot helper
└─ jobs/
   ├─ run_experiment.py           # orchestrates one D₀ ⊕ Dₖ experiment (multi sizes)
   ├─ train.py                    # PEFT/QLoRA SFT
   ├─ eval.py                     # metrics (loss/F1/Exact-Match)
   └─ scaling.py                  # fit & predict scaling law
```

---

## 2) Configuration & Env

**Space Settings → Secrets/Variables (already done for step 2, list here for reference):**

* `SERVICE_HF_TOKEN` (secret, write‑scoped; used to create/push results datasets)
* `RESULTS_REPO` (optional, like `your-org/curation-results`; if absent, create on first run)
* `HF_HOME=/data/.huggingface` (variable) **if** Persistent Storage is enabled
* `PERSIST_DIR=/data` (variable) **if** Persistent Storage is enabled

**NOTE: RESULTS_REPO is absent now; Persistent Storage is NOT enabled yet.**

**Runtime assumptions:**

* Space uses **Gradio SDK**.
* Jobs will request a **GPU flavor** (e.g., `a10g-small`) for training; UI itself can run on CPU.

**Currently the Space Hardware is ZeroGPU.**


---

## 3) Dependencies

`requirements.txt`

```
gradio>=5
huggingface_hub>=0.25
datasets>=2.20
transformers>=4.44
peft>=0.13
trl>=0.9
evaluate>=0.4
scikit-learn>=1.5
numpy>=1.26
pandas>=2.2
matplotlib>=3.8
```

---

## 4) Candidate Datasets Catalog

`catalog/candidates.json` (minimal starter; adjust to your domain)

```json
[
  {
    "id": "glue/sst2",
    "task": "classification",
    "license": "open",
    "size_hint": "67k",
    "columns": {"text": "sentence", "label": "label"},
    "labels": ["negative","positive"]
  },
  {
    "id": "ag_news",
    "task": "classification",
    "license": "cc-by-3.0",
    "size_hint": "120k",
    "columns": {"text": "text", "label": "label"},
    "labels": ["World","Sports","Business","Sci/Tech"]
  },
  {
    "id": "squad",
    "task": "qa",
    "license": "cc-by-sa-4.0",
    "size_hint": "100k",
    "columns": {"question": "question", "context": "context", "answers": "answers"}
  }
]
```

> For MVP, support **classification** and **extractive QA**. The `columns` mapping lets us normalize heterogeneous datasets without complex UI.

---

## 5) UI — `app.py` (Gradio)

### 5.1 Features

* **LoginButton** (OAuth) → captures `gr.OAuthProfile` and `gr.OAuthToken`.
* **D₀ input**: either upload files (`.jsonl/.csv/.parquet/.zip`) or provide a **Hub dataset id**.
* **Model** dropdown: start with `meta-llama/Llama-3.1-8B-Instruct`.
* **Task** selector (classification or QA). (MVP: single task per run.)
* **Benchmark/test set**: upload small test data or provide Hub split.
* **Metrics** checkboxes: `loss`, `f1`, `exact_match` (show `exact_match` only for QA).
* **Candidate datasets**: multiselect from `candidates.json`.
* **Run experiments** button: submits **one Job per selected Dₖ**.
* **Jobs table**: ID, Dₖ, status, logs link, artifacts link.
* **Results view**: scaling plot + ranked table when jobs finish.

### 5.2 Implementation Sketch

* Parse OAuth token; we’ll prefer the user token for **reading gated models**, but use `SERVICE_HF_TOKEN` for **writing** artifacts.
* If user **uploads D₀**, compress if needed and push to a **private dataset repo** via `utils/hub.ensure_uploaded_dataset(...)`.
* Submit a **Job** per Dₖ with:

  * command: `python jobs/run_experiment.py --model ... --d0 ... --dk ... --task ... --metrics ... --results_repo ...`
  * `flavor="a10g-small"` (configurable)
  * `timeout` (e.g., 7200 seconds)
  * `env`: `HF_TOKEN` (read), `SERVICE_HF_TOKEN` (write), plus `RESULTS_REPO` if set.
* Store job metadata in a `gr.State` list; start a **poller** (every ~10–15s) to refresh status via `huggingface_hub.inspect_job(...)`.
* When a job completes, show a link to its artifacts (scaling plot, metrics JSON) and update the results table.

**Acceptance criteria**

* Launching a run queues N jobs (N = number of selected Dₖ).
* Status column transitions through “queued/running/completed/failed”.
* Clicking an artifacts link opens an image/json from results repo (or Space storage).

---

## 6) Hub Utilities — `utils/hub.py`

### Functions to implement

* `ensure_uploaded_dataset(upload_files, d0_dataset_id, user_token) -> str`

  * If `d0_dataset_id` is provided, return it.
  * Else create a **private dataset repo** under your org (e.g., `your-org/curation-upload-<uuid>`), upload files/folder, and return repo id.
* `ensure_results_repo(service_token, results_repo_env) -> str`

  * If `RESULTS_REPO` is set, ensure it exists; else create `your-org/curation-results`.
* `push_artifacts(repo_id, local_dir, subdir) -> None`

  * Upload a local folder (e.g., `artifacts/<job-id>/...`) to `repo_id/subdir`.

**Acceptance criteria**

* Uploading a small CSV/JSONL creates a private dataset and returns a valid repo id.
* Pushing artifacts creates/updates files in the results repo with versioned commits.

---

## 7) Data Helpers — `utils/data.py`

### Responsibilities

* Load D₀ and Dₖ from the Hub (and optional **test set**).
* Normalize columns using the `columns` mapping from `candidates.json` or a provided override.
* Build **mixtures** of D₀ ⊕ Dₖ at multiple sizes (e.g., `{10k, 20k, 40k}` examples).
* For **classification**: expect `{"text": str, "label": int}` after normalization.
  For **QA**: expect `{"question": str, "context": str, "answers": {"text":[...], "answer_start":[...]}}`.

### API

```python
def load_dataset_normalized(repo_or_id, task, columns_map=None, split="train"):
    """Return a datasets.Dataset with normalized columns for the given task."""
    ...

def build_mixtures(d0_ds, dk_ds, sizes=[10_000, 20_000, 40_000], d0_ratio=0.5, seed=42):
    """Return dict: size -> datasets.Dataset of mixed examples (shuffled, repeat/trim as needed)."""

def load_benchmark(repo_or_id_or_path, task, split="validation"):
    """Return a small test set normalized for the chosen task."""
```

**Acceptance criteria**

* Given a known dataset id, `load_dataset_normalized(...)` returns columns as specified.
* `build_mixtures(...)` returns ≥2 sizes with the requested counts.

---

## 8) Plotting Helper — `utils/plotting.py`

### API

```python
def plot_scaling(sizes, y_values, y_label, out_path):
    """Save a simple matplotlib PNG (log-x) with points + fitted curve if provided."""
```

* Use matplotlib; one figure per plot; do not enforce custom colors/styles.

**Acceptance criteria**

* Calling `plot_scaling(...)` produces a PNG saved to `out_path` without errors.

---

## 9) Training — `jobs/train.py` (PEFT/QLoRA SFT)

**NOTE: Currently the Space Hardware is ZeroGPU. For testing purpose, the training part can be replaced by extremely small models.**


### Responsibilities

* Load model + tokenizer (e.g., `meta-llama/Llama-3.1-8B-Instruct`).
* Apply LoRA (or QLoRA).
* Tokenize dataset and run short SFT.

### API (sketch)

```python
from transformers import AutoTokenizer, AutoModelForCausalLM, TrainingArguments
from peft import LoraConfig, get_peft_model
from trl import SFTTrainer

def train_peft(model_id, train_ds, output_dir, max_steps=500, lr=2e-4, lora_r=8):
    tok = AutoTokenizer.from_pretrained(model_id, use_fast=True)
    base = AutoModelForCausalLM.from_pretrained(model_id)
    peft_cfg = LoraConfig(r=lora_r, lora_alpha=16, lora_dropout=0.05, task_type="CAUSAL_LM")
    model = get_peft_model(base, peft_cfg)

    def format_example(ex):
        # classification: concatenate prompt; QA: question + context formatting
        # MVP: simple "<s>[INST] ... [/INST]" style or plain text target
        return {"text": ex["text"]}  # adjust per task

    # Tokenization & SFTTrainer; keep it simple for MVP
    tr_args = TrainingArguments(output_dir=output_dir, per_device_train_batch_size=4,
                                gradient_accumulation_steps=4, learning_rate=lr,
                                max_steps=max_steps, logging_steps=50, save_steps=0)
    trainer = SFTTrainer(model=model, tokenizer=tok, train_dataset=train_ds,
                         dataset_text_field="text", args=tr_args)
    trainer.train()
    # Save adapter only
    trainer.save_model(output_dir)
    return output_dir
```

**Acceptance criteria**

* On a tiny dataset (few hundred samples), training completes and saves an adapter folder.

---

## 10) Evaluation — `jobs/eval.py`

### Responsibilities

* Run evaluation for the selected task using the fine‑tuned adapter.
* For **classification**: compute `loss` (optional) and `f1`.
* For **QA**: compute `exact_match` (and `f1` if you want both).

### API (sketch)

```python
import evaluate
import numpy as np

def eval_classification(model_id_or_path, test_ds):
    # Use pipeline or model.generate + simple argmax classifier (MVP)
    # Better: a small classification head; MVP keeps it simple.
    f1 = evaluate.load("f1")
    preds, refs = ..., ...
    return {"f1": f1.compute(predictions=preds, references=refs)["f1"]}

def eval_qa(model_id_or_path, test_ds):
    exact = evaluate.load("exact_match")
    # MVP: heuristic span matching if using generative outputs;
    # or reuse baseline SQuAD eval if test_ds has 'answers'.
    em = exact.compute(predictions=preds, references=refs)["exact_match"]
    return {"exact_match": em}
```

> **Note:** For MVP, inference can be slow. Keep test sets **small** (e.g., 500–1,000 examples) and batch where possible.

**Acceptance criteria**

* For a toy dataset, returns a metrics dict with expected keys.

---

## 11) Scaling Law — `jobs/scaling.py`

### Responsibilities

* Fit a simple power‑law over points `(size → metric)`.
* For “higher‑is‑better” metrics, convert to a pseudo‑loss (e.g., `1 - score`) during fitting if desired.
* Produce a **prediction** at a user‑defined large‑scale target (e.g., `N* = 200k` examples).

### API (sketch)

```python
import numpy as np

def fit_powerlaw(sizes, scores, higher_is_better=True):
    sizes = np.asarray(sizes, float)
    y = np.asarray(scores, float)
    if higher_is_better:
        # Fit to (1 - score) ~ b * N^{-alpha}
        z = np.log(np.maximum(1e-9, 1 - y))
    else:
        # Direct loss scaling
        z = np.log(np.maximum(1e-9, y))
    x = np.log(sizes)
    k, c = np.polyfit(x, z, 1)         # z ≈ k*log N + c
    alpha = -k; b = np.exp(c)
    return {"alpha": float(alpha), "b": float(b)}

def predict_powerlaw(size, fit_params, higher_is_better=True):
    alpha, b = fit_params["alpha"], fit_params["b"]
    if higher_is_better:
        loss_hat = b * (size ** (-alpha))
        return float(1 - loss_hat)
    return float(b * (size ** (-alpha)))
```

**Acceptance criteria**

* Given ≥2 points (prefer 3+), returns fit parameters and a plausible prediction.
* Combined with `utils/plotting.plot_scaling(...)`, writes a PNG with points + curve.

---

## 12) Experiment Orchestrator — `jobs/run_experiment.py`

### Responsibilities

* Parse CLI args: `--model`, `--task`, `--d0`, `--dk`, `--metrics ...`, `--sizes 10000 20000`, `--target_size 200000`, `--results_repo <id>`, `--job_id <uuid>`.
* Create working dirs: `artifacts/<job_id>/`.
* Load datasets (D₀, Dₖ), build mixtures for requested sizes.
* For each size:

  1. run short **train** (adapter saved under `artifacts/<job_id>/adapters/size-<N>`),
  2. run **eval** on the benchmark set → collect metrics.
* Fit **scaling** across sizes; produce:

  * `metrics.json` (per‑size metrics, fit params, predicted large‑scale performance),
  * `scaling.png` (plot).
* Push `artifacts/<job_id>/` to `results_repo` under `experiments/<user>/<job_id>/...` using `utils/hub.push_artifacts(...)`.
* Print a final JSON line to stdout with the artifacts path (UI can parse logs if needed).

### CLI Skeleton

```python
import argparse, json, os, uuid
from utils import hub, data, plotting
from jobs import train, eval as evalm, scaling

def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--model", required=True)
    ap.add_argument("--task", choices=["classification","qa"], required=True)
    ap.add_argument("--d0", required=True)
    ap.add_argument("--dk", required=True)
    ap.add_argument("--metrics", nargs="+", default=["f1"])
    ap.add_argument("--sizes", nargs="+", type=int, default=[10000, 20000, 40000])
    ap.add_argument("--target_size", type=int, default=200000)
    ap.add_argument("--results_repo", default=os.getenv("RESULTS_REPO",""))
    ap.add_argument("--job_id", default=str(uuid.uuid4()))
    args = ap.parse_args()

    # Setup dirs
    out_dir = os.path.abspath(os.path.join("artifacts", args.job_id))
    os.makedirs(out_dir, exist_ok=True)

    # Load datasets
    d0 = data.load_dataset_normalized(args.d0, args.task)
    dk = data.load_dataset_normalized(args.dk, args.task)
    test = data.load_benchmark(args.d0, args.task, split="validation")  # MVP: reuse D₀ val if none provided

    # Build mixtures & run train/eval
    per_size = []
    for N in args.sizes:
        mix = data.build_mixtures(d0, dk, sizes=[N])[N]
        adapter_dir = os.path.join(out_dir, f"adapter_size_{N}")
        train.train_peft(args.model, mix, adapter_dir, max_steps=300)  # MVP: few steps
        metrics = {}
        if args.task == "classification":
            metrics.update(evalm.eval_classification(adapter_dir, test))
        else:
            metrics.update(evalm.eval_qa(adapter_dir, test))
        per_size.append({"size": N, "metrics": metrics})

    # Fit scaling on the primary metric
    key = "exact_match" if args.task == "qa" else "f1"
    sizes = [r["size"] for r in per_size]
    scores = [r["metrics"][key] for r in per_size]
    fit = scaling.fit_powerlaw(sizes, scores, higher_is_better=True)
    pred = scaling.predict_powerlaw(args.target_size, fit, higher_is_better=True)

    # Write artifacts
    mpath = os.path.join(out_dir, "metrics.json")
    with open(mpath, "w") as f:
        json.dump({"runs": per_size, "fit": fit, "prediction": { "target_size": args.target_size, key: pred }}, f, indent=2)

    plotting.plot_scaling(sizes, scores, key, os.path.join(out_dir, "scaling.png"))

    # Push artifacts
    repo_id = hub.ensure_results_repo(os.getenv("SERVICE_HF_TOKEN"), args.results_repo)
    hub.push_artifacts(repo_id, out_dir, subdir=f"experiments/{args.job_id}")

    print(json.dumps({"status":"ok","artifacts_repo": repo_id, "path": f"experiments/{args.job_id}"}))

if __name__ == "__main__":
    main()
```

**Acceptance criteria**

* Running with tiny toy inputs creates `artifacts/<job_id>/` + pushes to results repo.
* `metrics.json` and `scaling.png` exist and look sensible.

---

## 13) Job Submission from UI — `app.py` (continued)

### Core actions

* **Submit**: for each selected Dₖ → call `huggingface_hub.run_job(...)` with:

  * `image`: CUDA‑capable (e.g., `pytorch/pytorch:2.6.0-cuda12.4-cudnn9-devel`)
  * `command`: `["python","jobs/run_experiment.py", "--model", model_id, "--task", task, "--d0", d0_repo, "--dk", dk_id, "--metrics", *metrics, "--sizes", *sizes, "--target_size", str(target_size), "--results_repo", results_repo_or_empty]`
  * `flavor`: `"a10g-small"`
  * `timeout`: e.g., `7200` (seconds)
  * `env`: `{"HF_TOKEN": user_token or SERVICE_HF_TOKEN, "SERVICE_HF_TOKEN": SERVICE_HF_TOKEN, "RESULTS_REPO": RESULTS_REPO}`

* **Poll**: keep a dict `{job_id: {dk, status, url, artifacts}}`; update via `inspect_job(job_id)`; for `completed`, set artifacts link to `hf://<results_repo>/experiments/<job_id>/`.

**Acceptance criteria**

* Submitting 2 Dₖ creates 2 jobs; both progress independently; artifacts link works.

---

## 14) Guardrails & Licensing

* **Gated models**: probe download with `hf_hub_download(model_id, filename="README.md", token=user_token)` to confirm access; if 401/403, show a clear message to accept the license on the model card.
* **Dataset licensing**: surface the `license` field from `candidates.json` next to each Dₖ; later fetch from Hub.
* **Uploads**: warn users that uploaded D₀ will be stored in a **private dataset** (repo id shown in UI); provide a “Delete my upload” note linking to the repo.
* **Resource limits**: cap sizes (`sizes=[5_000, 10_000]` for MVP), cap number of concurrent jobs per user (client‑side only for MVP).

---

## 15) Testing

### Local (CPU) sanity checks

* Use a very small subset (e.g., 200 examples) and `max_steps=10` to verify the end‑to‑end loop without a GPU.
* Mock `run_job(...)` (optional) to test UI job table logic.

### Space integration

* Create a private test Space results repo (e.g., `your-org/curation-results-test`).
* Submit a single Dₖ job and verify:

  * `artifacts/` created,
  * `metrics.json` contains per‑size metrics and prediction,
  * `scaling.png` renders,
  * artifacts are uploaded and visible from the UI link.

---

## 16) Definition of Done (DoD)

* A signed‑in user can:

  1. Provide **D₀** (upload or Hub id),
  2. Choose **model**, **task**, **metrics**, and ≥1 **Dₖ**,
  3. Click **Run** and see a job per Dₖ with live status,
  4. Open **artifacts** (plot + metrics),
  5. See a **ranked table** of Dₖ by the chosen primary metric,
  6. (Optional) Download `metrics.json`.

* All long work executes as **Jobs** (no HTTP timeouts).

* Artifacts persist in a results dataset or Space storage.

---

## 17) Nice‑to‑Have (post‑MVP)

* **Column mapping UI**: let users map their D₀ columns to `text/label` or `question/context/answers`.
* **Seed sweeps** and confidence intervals on scaling fit.
* **Hardware selector** and budget estimator.
* **vLLM/TGI** inference for faster eval.
* **Per‑user “My Experiments”** page (prefix `experiments/<username>/...`).

---

## 18) Task Checklist (assignable to your agent)

**A. Scaffolding**

* [ ] Add `requirements.txt`; ensure importable on the Space.
* [ ] Create folders: `catalog/`, `utils/`, `jobs/`.

**B. Catalog**

* [ ] Fill `catalog/candidates.json` (3–6 datasets), including `columns` mapping.

**C. Hub utilities (`utils/hub.py`)**

* [ ] `ensure_uploaded_dataset(...)`
* [ ] `ensure_results_repo(...)`
* [ ] `push_artifacts(...)`

**D. Data helpers (`utils/data.py`)**

* [ ] `load_dataset_normalized(...)` for classification + QA
* [ ] `build_mixtures(...)`
* [ ] `load_benchmark(...)`

**E. Plotting (`utils/plotting.py`)**

* [ ] `plot_scaling(...)`

**F. Jobs**

* [ ] `jobs/train.py` (PEFT SFT)
* [ ] `jobs/eval.py` (classification + QA)
* [ ] `jobs/scaling.py` (fit + predict)
* [ ] `jobs/run_experiment.py` (glue the above, produce artifacts, push)

**G. UI (`app.py`)**

* [ ] Build form (inputs, selectors, candidates list)
* [ ] Submit one job per Dₖ via `run_job(...)`
* [ ] Poll job status & render jobs table
* [ ] Artifacts viewer: link to results repo path
* [ ] Basic error messages (license issues, upload failures)

**H. Tests**

* [ ] Local micro‑run (CPU) with tiny sizes
* [ ] Space run on GPU flavor with one Dₖ
* [ ] Verify artifacts + plot + ranking table

---

## 19) Code Snippets to Start Implementation

### `app.py` — minimal UI skeleton (submit + poll)

```python
import os, json, time, gradio as gr
from huggingface_hub import run_job, inspect_job
from utils.hub import ensure_uploaded_dataset, ensure_results_repo

CANDIDATES = json.load(open("catalog/candidates.json"))

def submit(d0_files, d0_id, task, model, metrics, dk_list, sizes, target_size,
           profile: gr.OAuthProfile | None, oauth: gr.OAuthToken | None):
    user_token = getattr(oauth, "token", None)
    d0_repo = ensure_uploaded_dataset(d0_files, d0_id, user_token=user_token)
    results_repo = ensure_results_repo(os.getenv("SERVICE_HF_TOKEN"), os.getenv("RESULTS_REPO",""))
    jobs = []
    for dk in dk_list:
        cmd = ["python","jobs/run_experiment.py",
               "--model", model, "--task", task, "--d0", d0_repo, "--dk", dk,
               "--metrics", *metrics, "--sizes", *[str(s) for s in sizes],
               "--target_size", str(target_size), "--results_repo", results_repo]
        job = run_job(
            image="pytorch/pytorch:2.6.0-cuda12.4-cudnn9-devel",
            command=cmd,
            flavor="a10g-small",
            timeout=7200,
            env={"HF_TOKEN": user_token or os.getenv("SERVICE_HF_TOKEN"),
                 "SERVICE_HF_TOKEN": os.getenv("SERVICE_HF_TOKEN"),
                 "RESULTS_REPO": results_repo},
        )
        jobs.append({"id": job.id, "dk": dk, "url": job.url, "status": "queued", "artifacts": ""})
    return jobs

def poll(jobs_state):
    updated = []
    for j in jobs_state:
        info = inspect_job(j["id"])
        st = info.status  # "queued"/"running"/"completed"/"failed"
        art = j.get("artifacts","")
        # Heuristic: artifacts live in RESULTS_REPO/experiments/<job_id> (set by run_experiment.py)
        if st == "completed" and not art:
            art = f"{os.getenv('RESULTS_REPO','(repo)')}/experiments/{j['id']}"
        updated.append({**j, "status": st, "artifacts": art})
    return updated

with gr.Blocks() as demo:
    prof = gr.LoginButton()
    with gr.Row():
        d0_files = gr.UploadButton("Upload D₀ (.csv/.jsonl/.zip)", file_count="multiple")
        d0_id = gr.Textbox(label="or Hub dataset id (user/dataset)")
    task = gr.Radio(choices=["classification","qa"], value="classification", label="Task")
    model = gr.Dropdown(choices=["meta-llama/Llama-3.1-8B-Instruct"], label="Model")
    metrics = gr.CheckboxGroup(choices=["loss","f1","exact_match"], value=["f1"], label="Metrics")
    dk = gr.CheckboxGroup(choices=[c["id"] for c in CANDIDATES], label="Candidate datasets")
    sizes = gr.CheckboxGroup(choices=[5000,10000,20000], value=[5000,10000], label="Mixture sizes")
    target_size = gr.Number(value=200000, label="Target size for prediction")
    run_btn = gr.Button("Run experiments")

    jobs_state = gr.State([])
    jobs_table = gr.Dataframe(headers=["id","dk","status","url","artifacts"], datatype=["str","str","str","str","str"])

    run_btn.click(fn=submit,
                  inputs=[d0_files, d0_id, task, model, metrics, dk, sizes, target_size, gr.OAuthProfile, gr.OAuthToken],
                  outputs=jobs_state)

    gr.Button("Refresh status").click(fn=poll, inputs=jobs_state, outputs=jobs_state)

    def render_table(jobs):  # render as simple rows
        rows = [[j["id"], j["dk"], j["status"], j["url"], j["artifacts"]] for j in jobs]
        return rows
    jobs_state.change(fn=render_table, inputs=jobs_state, outputs=jobs_table)

    gr.Markdown("Open artifacts in the results repo once jobs complete.")

demo.queue().launch()
```

### `utils/hub.py` — upload & results

```python
import os, uuid, tempfile, shutil
from huggingface_hub import HfApi, create_repo, upload_file, upload_folder

def ensure_uploaded_dataset(upload_files, d0_dataset_id, user_token=None):
    if d0_dataset_id:
        return d0_dataset_id
    if not upload_files:  # nothing uploaded
        raise ValueError("Please upload D₀ or provide a Hub dataset id.")
    api = HfApi(token=os.getenv("SERVICE_HF_TOKEN"))
    repo_id = f"{os.getenv('HF_ORG','your-org')}/curation-upload-{uuid.uuid4().hex[:8]}"
    create_repo(repo_id, repo_type="dataset", private=True, exist_ok=True, token=os.getenv("SERVICE_HF_TOKEN"))

    with tempfile.TemporaryDirectory() as tmp:
        # Gradio returns a list of tempfiles; copy them into a folder
        for f in upload_files:
            dst = os.path.join(tmp, os.path.basename(getattr(f,"name", "file")))
            shutil.copyfile(f.name if hasattr(f,"name") else f, dst)
        upload_folder(folder_path=tmp, repo_id=repo_id, repo_type="dataset", token=os.getenv("SERVICE_HF_TOKEN"))
    return repo_id

def ensure_results_repo(service_token, results_repo_env):
    api = HfApi(token=service_token)
    if results_repo_env:
        parts = results_repo_env.split("/")
        if len(parts) == 2:
            create_repo(results_repo_env, repo_type="dataset", private=True, exist_ok=True, token=service_token)
            return results_repo_env
    repo_id = f"{os.getenv('HF_ORG','your-org')}/curation-results"
    create_repo(repo_id, repo_type="dataset", private=True, exist_ok=True, token=service_token)
    return repo_id

def push_artifacts(repo_id, local_dir, subdir=""):
    path_in_repo = subdir.strip("/")
    upload_folder(folder_path=local_dir, repo_id=repo_id, repo_type="dataset",
                  path_in_repo=path_in_repo if path_in_repo else None,
                  token=os.getenv("SERVICE_HF_TOKEN"))
```