Trim the card; the builder writes the JSONL prompts only

README.md

@@ -17,20 +17,15 @@ configs:
 
 # Quant fidelity eval (public sources)
 
-This dataset holds prompts for measuring how closely a quantized model reproduces its full-precision original. All 240
-prompts come from seven public datasets.
+240 prompts from seven public datasets for measuring how closely a quantized model reproduces its original, as
+per-token KL divergence.
 
-The measurement is per-token KL divergence between the original model and the quantized one. It is computed on two
-kinds of text:
+- **`on_policy`** (140): the original model continues each prompt greedily for 512 tokens; each quant is scored on
+  that text.
+- **`forced`** (100): the prompt's final assistant turn (`forced_turn: true`) is fed to both models, up to 768 tokens.
 
-- **`on_policy`** (140 prompts): the original model continues each prompt greedily, here for 512 tokens. Each quant is
-  then scored on that same text.
-- **`forced`** (100 prompts): the dataset's own assistant turn is fed to both models (teacher forcing), here up to 768
-  tokens. The turns include reasoning, code edits and tool calls.
-
-Prompts are stored as chat messages and are rendered with the evaluated model's own chat template when the reference
-is frozen. A prompt with `forced_turn: true` ends in the assistant turn to force; every other prompt ends where the
-assistant should reply.
+Each line holds `id`, `category`, `messages` and `forced_turn`. Render the messages with the evaluated model's chat
+template.
 
 ## Sources
 
@@ -44,60 +39,17 @@ assistant should reply.
 | openhands | [SWE-Gym/OpenHands-SFT-Trajectories](https://huggingface.co/datasets/SWE-Gym/OpenHands-SFT-Trajectories), `train.success.oss` | MIT | 20 | 20 |
 | glaive_fc | [glaiveai/glaive-function-calling-v2](https://huggingface.co/datasets/glaiveai/glaive-function-calling-v2) | Apache-2.0 | 20 | 15 |
 
-The prompt text belongs to these datasets and is used here for evaluation only. Their licences and attribution terms
-apply.
-
-## Selection
-
-Selection is deterministic, so anyone can rebuild the set with `build_public.py`.
-
-- **Pinned revisions.** Each dataset is read at the revision recorded in `public_manifest.json`.
-- **Order.** Rows are read in a seeded streaming shuffle (seed 0, buffer 5000). OpenHands is read in file order.
-- **Quotas.** Rows fill the on-policy quota first, then the forced quota. Each row is used at most once.
-- **Cut point.** The turn to cut at, or to force, is `sha256(source, row id, set) mod (number of eligible turns)`.
-- **Eligible turns.** At most 24,000 characters of context may come before the cut. A forced turn must be at least
-  800 characters for chat sources, 400 for the agent trajectories and 150 for function calling.
-- **Format conversions.**
-  - SWE-agent: `ai` turns become assistant messages.
-  - Glaive: `FUNCTION RESPONSE` turns become user messages prefixed "Function response:".
-
-`public_manifest.json` lists, for every prompt, the dataset, revision, split, row id and turn index.
-
-## Files
-
-| File | Contents |
-| --- | --- |
-| `public_on_prompts.jsonl`, `public_off_prompts.jsonl` | One prompt per line: `id`, `category`, `messages` and, for forced prompts, `forced_turn` |
-| `public_on_prompts.json`, `public_off_prompts.json` | The same prompts in the prompt-file format the autoadd v2 reference freezer reads |
-| `public_manifest.json` | How each prompt was selected |
-| `build_public.py` | The builder |
+The prompt text belongs to these datasets; their licences and attribution terms apply.
 
 ## Usage
 
-1. **Freeze the references.** Freeze the original model's outputs once, for example with autoadd v2:
-
-   ```
-   autoadd_ref.py --hf-model <bf16 checkpoint> --out-dir <dir> \
-     --set public_on=public_on_prompts.json:512:floor=0 \
-     --set public_off=public_off_prompts.json:768:floor=0
-   ```
-
-2. **Score each quant.** Score with autoquant v2:
-
-   ```
-   python -m autoquant.eval.measure measure --engine gguf --model <quant.gguf> \
-     --goldens <dir>/reference_public_on.json --logprob-dir <dir>/ref_logprobs_public_on \
-     --off-goldens <dir>/reference_public_off.json --off-logprob-dir <dir>/ref_logprobs_public_off --out result.json
-   ```
-
-3. **Compare.** Report the pooled mean, p50, p99 and max KL. Bootstrap over prompts, not tokens: tokens within one
-   prompt are strongly correlated.
+1. Freeze the original model's log-probabilities at every continuation position (greedy for `on_policy`, the given
+   turn for `forced`).
+2. Run each quant on the same tokens and compute KL(original ‖ quant) per token.
+3. Report the pooled mean, p50, p99 and max, and bootstrap over prompts rather than tokens.
 
-## Caveats
+## Rebuilding
 
-- **Frozen references are not included.** References depend on the model and the hardware, and their log-prob rows are
-  tens of GB. Freeze your own.
-- **The base model may have seen these datasets in training.** That does not affect a comparison between quants of the
-  same model, because KL measures only how faithfully each quant reproduces the original.
-- **UltraChat KL is unusually high** for every quant we have measured, about 10× the other sources. It is still being
-  investigated, so report results both with and without it.
+`python build_public.py <out dir>` (needs `datasets` and `huggingface_hub`) rebuilds the set exactly: pinned
+revisions, a seeded shuffle and hash-chosen cut points. `public_manifest.json` records the source row and turn of every
+prompt.

build_public.py

@@ -1,5 +1,12 @@
-"""Public goldens prompt sets for quant fidelity (no internal data). Deterministic: pinned revisions, seeded
-streaming shuffle, hash-chosen cut points. Writes autoadd-format prompt files + a publishable manifest."""
+"""Builds the quant-fidelity prompt sets from public datasets.
+
+usage: python build_public.py <out dir>
+
+Selection is deterministic: each dataset is read at a pinned revision in a seeded streaming shuffle, rows fill the
+on-policy quota and then the forced quota, and the turn to cut at is chosen by hash. Writes public_on_prompts.jsonl,
+public_off_prompts.jsonl and public_manifest.json.
+"""
+
 import hashlib, json, re, sys
 from datasets import load_dataset
 OUT = sys.argv[1]; SEED = 0; MAX_PREFIX = 24000
@@ -45,7 +52,7 @@ def norm_glaive(r):
         elif tag == "ASSISTANT:": msgs.append({"role": "assistant", "content": body})
         else: msgs.append({"role": "user", "content": "Function response: " + body})
     return hashlib.sha256((r["system"] + r["chat"]).encode()).hexdigest()[:16], msgs
-SOURCES = [  # name, repo, revision, split, normaliser, greedy quota, forced quota, min forced chars, shuffle
+SOURCES = [  # name, repo, revision, split, normaliser, on-policy quota, forced quota, min forced chars, shuffle
  ("ultrachat", "HuggingFaceH4/ultrachat_200k", "8049631c40", "test_sft", norm_ultrachat, 20, 15, 800, True),
  ("openhermes", "teknium/OpenHermes-2.5", "b820378210", "train", norm_hermes, 20, 15, 800, True),
  ("tulu3", "allenai/tulu-3-sft-mixture", "b14afda60f", "train", norm_tulu, 25, 15, 800, True),
@@ -81,11 +88,12 @@ for name, repo, rev, split, norm, qg, qf, minf, shuf in SOURCES:
             off.append({"id": f"{name}-{nf:02d}", "category": name, "messages": msgs[:k + 1], "forced_turn": True})
             manifest.append({"set": "public_off", "id": f"{name}-{nf:02d}", "dataset": repo, "revision": full_rev, "split": split, "row_id": rid, "forced_message": k}); nf += 1
     print(f"{name}: greedy {ng}/{qg} forced {nf}/{qf} (rows read {seen})", flush=True)
-for fname, ents, what in (("public_on_prompts.json", on, "greedy BF16 continuation, 512 tokens"),
-                          ("public_off_prompts.json", off, "forced dataset assistant turn, up to 768 tokens")):
-    json.dump({"comment": f"public quant-fidelity set: {what}. Sources and selection in public_manifest.json.",
-               "key_tokenizer": True, "prompts": ents}, open(f"{OUT}/{fname}", "w"), ensure_ascii=False, indent=1, sort_keys=True)
+for fname, prompts in (("public_on_prompts.jsonl", on), ("public_off_prompts.jsonl", off)):
+    with open(f"{OUT}/{fname}", "w") as handle:
+        for prompt in prompts:
+            handle.write(json.dumps({**prompt, "forced_turn": prompt.get("forced_turn", False)}, ensure_ascii=False) + "\n")
 json.dump({"seed": SEED, "max_prefix_chars": MAX_PREFIX, "selection": "seeded streaming shuffle (buffer 5000) except OpenHands; "
-           "rows in order fill the greedy quota first, then forced; cut point = sha256(source, row id, set) mod candidates; Tulu WildChat rows excluded",
+           "rows in order fill the on-policy quota first, then forced; cut point = sha256(source, row id, set) mod candidates; "
+           "Tulu WildChat rows excluded",
            "prompts": manifest}, open(f"{OUT}/public_manifest.json", "w"), indent=1)
 print("on", len(on), "off", len(off))

public_manifest.json

@@ -1,7 +1,7 @@
 {
  "seed": 0,
  "max_prefix_chars": 24000,
- "selection": "seeded streaming shuffle (buffer 5000) except OpenHands; rows in order fill the greedy quota first, then forced; cut point = sha256(source, row id, set) mod candidates; Tulu WildChat rows excluded",
+ "selection": "seeded streaming shuffle (buffer 5000) except OpenHands; rows in order fill the on-policy quota first, then forced; cut point = sha256(source, row id, set) mod candidates; Tulu WildChat rows excluded",
  "prompts": [
   {
    "set": "public_on",