mayflowergmbh/boldt-dc-1b-orpo-onpolicy-de

Boldt-DC-1B On-Policy ORPO (German)

A 56,413-row German preference dataset in which the rejected field is the actual output of the mayflowergmbh/boldt-dc-1b-german-it-16k SFT model under greedy decoding, rather than an off-the-shelf "alternative" answer.

The intent is on-policy ORPO/DPO: training against a model's own current failure modes typically gives a much sharper gradient signal than training against a generic weak baseline, because every preference pair documents a concrete gap between what this specific model does today (rejected) and what we want it to do (chosen).

Quick start

from datasets import load_dataset
ds = load_dataset("mayflowergmbh/boldt-dc-1b-orpo-onpolicy-de", split="train")
print(ds[0]["prompt"][-300:])
print("CHOSEN:",   ds[0]["chosen"][:200])
print("REJECTED:", ds[0]["rejected"][:200])

Drop straight into TRL's ORPOTrainer / DPOTrainer (the prompt field is already formatted with the Boldt chat tokens <|system|> / <|user|> / <|assistant|> and primed with a trailing <|assistant|>\n, so no chat-template application is required).

Schema

Column	Type	Description
prompt	string	Full conversation history + system prompt, formatted with Boldt's chat tokens, ending with `<
chosen	string	The preferred assistant response (from the source dataset)
rejected	string	Greedy continuation of prompt produced by mayflowergmbh/boldt-dc-1b-german-it-16k
original_rejected	string	The off-the-shelf rejected text from the source dataset, kept for provenance/diffing
source	string	"orpo-dpo-mix-40k-de" or "intel-orca-dpo-de"
row_id	string	24-char SHA-256 of source + "::" + prompt, for stable dedup and resumable builds
prompt_tok_len	int	Token count of the prompt (Boldt tokenizer)
chosen_tok_len	int	Whitespace word count of chosen
rejected_tok_len	int	Whitespace word count of rejected
is_refusal	bool	Flagged by a small German+English refusal pattern. Refusals are not dropped — when chosen is a real answer, a refusal in rejected is valid contrastive signal
gen_model	string	Generation model identifier
gen_engine	string	"hf" (transformers + manual batched generation)
gen_seed	int	RNG seed
gen_temperature	float	0.0 (greedy)
timestamp_utc	string	When the row was generated

How it was built

1. Source unification. Two German preference datasets were merged and unified into a (prompt, chosen, original_rejected) shape:

   Source	Rows ingested	Rows kept after unify + cross-source dedup
   johannhartmann/orpo-dpo-mix-40k-llama3-de	44,240	43,947
   mayflowergmbh/intel_orca_dpo_pairs_de	12,855	12,466

   The prompt string is built with the exact same helper (build_orpo_prompt in boldt_it.dataset_formatting) that ORPO training uses, so the model sees byte-identical conditioning at training time and at the moment its rejected was generated.

2. Generation. mayflowergmbh/boldt-dc-1b-german-it-16k was loaded in bf16 on a single RTX A6000. For each prompt, the model emitted a greedy continuation with do_sample=False, max_new_tokens=384, max_seq_length=4096. Generation was batched with padding_side="left" and length-bucketed within shuffled 256-row chunks; throughput was ~2.78 prompts/s end-to-end (≈5.6 h total).

3. Filtering. Post-generation drops:

   Reason	Count
   empty rejected	12
   too short (< 20 chars)	29
   equal to chosen	0
   near-duplicate (5-gram Jaccard ≥ 0.92)	0
   prompt would have been truncated (> 3,584 tokens)	39
   kept	56,413

   No equal_to_chosen and no jaccard_dup survived → there is real gradient signal in every published row.

Length distributions

	chosen (words)	rejected (words)
p10	30	108
median	216	237
mean	235	218
p90	457	294
p99	674	329
max	1,561	384

rejected capped at 384 tokens by the generation budget; chosen tail extends because the source datasets contain long-form answers. For ORPO this is fine — the contrast happens primarily in the first few hundred tokens.

Refusal rate on rejected: 1.2 % (flagged in is_refusal).

Intended use

• ORPO/DPO/IPO/SimPO post-training of mayflowergmbh/boldt-dc-1b-german-it-16k and close descendants. The on-policy property is specific to this base model; using this dataset to train a different model gives you off-policy data again.
• Diagnostic analysis of failure modes in the SFT-only Boldt model: every rejected row is a concrete example of where the model currently underperforms.

Limitations and considerations

• Quality of rejected ≠ minimum-quality reference. Some rejected outputs are actually reasonable answers that just differ stylistically from chosen. ORPO will still learn from these, but the gradient strength varies per row.
• chosen quality inherits from source datasets. orpo-dpo-mix-40k is a community-curated mix; intel_orca_dpo_pairs_de is a German translation of Intel's Orca DPO pairs. Neither has been hand-verified for factual accuracy.
• Repetition / hallucination patterns in rejected are characteristic of 1.25 B-parameter Llama-family SFT-only behavior. Don't assume these patterns will transfer to a different scale.
• Licensing. This dataset is released under Apache 2.0, but the upstream license of johannhartmann/orpo-dpo-mix-40k-llama3-de and mayflowergmbh/intel_orca_dpo_pairs_de (which seed the prompts and chosen answers) should be reviewed before commercial use.

Provenance

• Generation script: scripts/16_build_orpo_onpolicy_de.py in the Boldt training repo
• Build report: reports/orpo_on_policy_build.json
• Spot-check sample: reports/orpo_on_policy_spotcheck.md (20 rows of prompt / chosen / rejected / original_rejected)
• Engine: transformers 5.5.0 + torch 2.10.0+cu128, bf16, NVIDIA RTX A6000

Citation

@misc{boldt-onpolicy-orpo-de-2026,
  title  = {Boldt-DC-1B On-Policy ORPO (German)},
  author = {Mayflower GmbH},
  year   = {2026},
  url    = {https://huggingface.co/datasets/mayflowergmbh/boldt-dc-1b-orpo-onpolicy-de}
}