Firworks/no_robots-gemmafy

No Robots — Gemmafy

Every final assistant turn in HuggingFaceH4/no_robots regenerated by Gemma 4 31B-it, with the original human-written answer kept alongside for comparison.

Why: No Robots is a well-known set of 10,000 human-written prompts across ten everyday task categories. Pairing each prompt with both a human answer and Gemma's answer captures Gemma's voice — formatting habits, length, tone, persona handling — on identical inputs, for voice/style-transfer research.

Nothing in Gemma's replies was fact-checked or edited. See Limitations.

Part of the Gemmafy series. A Gemmafy takes a well-known dataset, regenerates its responses with Gemma 4 31B-it, and keeps the originals alongside for comparison. Browse the series with the gemmafy tag.

Usage

from datasets import load_dataset

ds = load_dataset("Firworks/no_robots-gemmafy")

# Optional: drop the 13 rows with a quality flag (see "Flags" below)
flags = ["truncated", "empty", "refusal", "repetitive", "thinking_leak", "self_correction"]
clean = ds.filter(lambda r: not any(r[f] for f in flags))

# Optional: keep only conversations where every assistant turn is Gemma's
gemma_only = ds.filter(lambda r: not r["is_multiturn"])

Contents

split	rows	source split
train	9,500	train
test	500	test

The test split is a held-out evaluation set. Do not train on it.

category	train	test	mean Gemma tokens (train)
Generation	4,346	214	434
Open QA	1,182	58	343
Brainstorm	1,060	60	668
Chat	796	54	113
Rewrite	625	35	396
Summarize	395	25	115
Coding	334	16	697
Classify	334	16	136
Closed QA	245	15	56
Extract	183	7	75

Gemma's replies average ~388 completion tokens (median ~320); 3.88M completion tokens in total.

Columns

column	type	description
prompt_id	string	From No Robots. Not unique: two train rows (Chat, same system prompt) share one. Use source_row_idx to join.
source_row_idx	int	Row index in the corresponding No Robots split.
category	string	No Robots category.
prompt	string	No Robots prompt field (for system-prompt rows, this is the system prompt).
messages	list of {role, content}	Full conversation with the final assistant turn replaced by Gemma's reply. Ready for SFT.
original_response	string	The human-written final assistant turn from No Robots.
gemma_response	string	Gemma's reply (identical to messages[-1].content).
is_multiturn	bool	Conversation has more than one assistant turn. Earlier assistant turns are human-written.
has_system_prompt	bool	Conversation starts with a system message (a persona).
finish_reason	string	stop or length.
completion_tokens	int	Gemma completion tokens for this reply.
truncated, empty, refusal, repetitive, thinking_leak, self_correction	bool	Quality flags — see Flags.
role_anomaly	bool	Source-data quirk — see Multi-turn rows.

No rows were removed; problem rows are flagged instead.

Generation details

Model	google/gemma-4-31B-it, served from the pre-quantized checkpoint RedHatAI/gemma-4-31B-it-FP8-Dynamic (FP8 dynamic weights/activations), fp8 KV cache
Chat template	The checkpoint's bundled template, identical to Google's gemma-4-31B-it template
Engine	vLLM 0.32.1rc1.dev197+g8e2beaa3b, OpenAI-compatible chat completions API
Speculative decoding	MTP: {"method": "mtp", "model": "google/gemma-4-31B-it-assistant", "num_speculative_tokens": 5}
max-model-len	262,144
Sampling	temperature=1.0, top_p=0.95, top_k=64 — Google's recommended settings from the Gemma 4 model card
max_tokens	8,192
Thinking	Disabled per request with chat_template_kwargs={"enable_thinking": false}. The template then pre-fills an empty thought block, as the model card describes. No reply contains reasoning text or channel markup, and vLLM's reasoning parser returned no reasoning content for any row.
Seeds	None
Hardware	NVIDIA DGX Spark, 64 concurrent requests, steady-state ≈ 480 completion tok/s
Generated	2026-11-01 (UTC)

System role. Gemma 4 supports the system role natively. The template renders a leading system message as its own system turn (surrounding whitespace trimmed); it is not folded into the first user turn. Persona system prompts were sent exactly as written in No Robots.

Reproducibility. The scripts in scripts/ regenerate the dataset with the same procedure. Outputs are not bit-reproducible: no seeds were set, and vLLM's batched, speculative decoding is not deterministic across batch compositions. A rerun gives replies from the same distribution, not the same text.

Request construction

For each No Robots row, messages up to but not including the final assistant turn were sent; Gemma's reply replaces that turn. Every source row ends with an assistant turn, so none were skipped.

Multi-turn rows

849 rows (795 train, 54 test; all in the Chat category, all with a persona system prompt) have more than one assistant turn. Only the final assistant turn is Gemma's; the earlier assistant turns are still the human-written originals. These rows have is_multiturn = true. If you want purely Gemma-voiced conversations, filter them out or use them only for final-turn training.

role_anomaly (15 train rows): in the source data, these conversations end with two consecutive assistant turns. In most of the ones we read, the second-to-last "assistant" turn is clearly the user speaking, mislabeled in No Robots. We did not correct the source; the context was sent as-is (ending in an assistant turn) and Gemma's reply replaced the final turn. Gemma generally treated the mislabeled turn as the user's.

Flags

Flags are computed by the heuristics in scripts/flags.py. Every row a heuristic flagged was then read by hand, and the one false positive was cleared (listed in MANUAL_FALSE_POSITIVES in scripts/build_dataset.py), so every flag in the released data marks a real problem. The "heuristic hits" column shows what the patterns caught before review.

flag	definition	train	test	heuristic hits → confirmed
truncated	finish_reason == "length" (hit the 8,192-token limit)	1	0	1 → 1: a degenerate loop
empty	Blank or whitespace-only reply	0	0	—
refusal	Decline phrase (e.g. "I can't help with", "I must decline", "As an AI, I can't") in the first 300 characters; disclaimers before a full answer ("While I cannot…, …") excluded	1	0	2 → 1 (see below)
repetitive	≥ 80 words and > 30% of word 10-grams are repeats	3	0	3 → 3: loops or list repetition
thinking_leak	Reasoning/template markup in the reply (<|channel>, <channel|>, <|think|>, <think>, <|turn>…), a reply starting with thought\n, or a non-empty reasoning field	0	0	—
self_correction	Gemma visibly corrects itself mid-reply ("Wait, let me redo…", "(Correction: …)", "Repeated in thought, correcting below")	11	1	12 → 12

13 rows (12 train, 1 test) have at least one of these flags.

Refusals are almost absent. No Robots contains very few prompts a model would decline, and the flag is a phrase match, so it is mostly noise here. To estimate recall, we hand-read all 76 replies containing any broad decline-like phrase ("I'm sorry", "I can't", "I'm unable", "as an AI", …) near the start. Nearly all were in-story dialogue, drafted messages ("I am unable to work today"), persona banter, or capability caveats ("I can't see your code"). The only real decline is an in-character one: a "sad and threat-averse" persona refusing to recommend firecrackers. It is flagged. The heuristic's other hit was a false positive (a dad-joke persona answering "Forget it." with "I can't do that! I have a terrible memory.") and was cleared.

self_correction was added after reviewing the run. With thinking disabled, Gemma occasionally does its reasoning in the visible answer: it produces a list, notices a constraint violation or duplicate, and redoes it. It mostly happens on constrained list prompts (word games, "one per decade", alphabetical lists). These replies are usually also partly wrong. The pattern deliberately excludes "Corrected code:" / "Corrected version:", which are legitimate fixes of the user's input. Recall is unknown beyond the candidates we reviewed.

Legitimate repetition (song choruses, "repeat the previous sentence and add a word" prompts) reaches about 0.29 on the n-gram ratio and is not flagged.

Limitations

• Not fact-checked. Gemma's answers on factual (Open QA, Closed QA) and coding prompts were not verified and are sometimes wrong. Fine-tuning on messages teaches the content along with the style. Conversely, some No Robots originals are from 2023 and may themselves be outdated.
• Thinking off, temperature 1.0. This is Gemma with Google's recommended sampling and without its reasoning mode, which is not its most accurate configuration. The self_correction rows show the cost.
• Mixed authorship in multi-turn rows. In is_multiturn rows, only the last assistant turn is Gemma's.
• Heuristic flags. The flags are pattern matches. Every hit was hand-checked, so precision is effectively 100%, but recall is not guaranteed: problem rows the patterns don't match are unflagged.
• FP8 checkpoint. Generated with an FP8-quantized checkpoint, not the original BF16 weights. Small differences from BF16 outputs are expected.
• License. Non-commercial (CC BY-NC 4.0), inherited from No Robots. Gemma 4 is Apache 2.0 and adds no further restrictions.

Scripts

scripts/ contains everything used to build this dataset:

• generate.py: async, resumable generation against a vLLM OpenAI-compatible server. Rows are keyed by (split, row_idx); transient errors are retried with backoff.
• flags.py: the flag heuristics.
• build_dataset.py: joins generations with No Robots and writes the Parquet splits.
• env.example.sh: environment variables generate.py expects.

Source dataset and citation

Prompts, conversations and original_response come from HuggingFaceH4/no_robots (CC BY-NC 4.0). Please cite it:

@misc{no_robots,
  author = {Nazneen Rajani and Lewis Tunstall and Edward Beeching and Nathan Lambert and Alexander M. Rush and Thomas Wolf},
  title = {No Robots},
  year = {2023},
  publisher = {Hugging Face},
  journal = {Hugging Face repository},
  howpublished = {\url{https://huggingface.co/datasets/HuggingFaceH4/no_robots}}
}