genrec-dataset

Phase-2 reasoning data for SIDReasoner. The reasoning_path column holds chain-of-thought traces generated with Azure OpenAI GPT-5.4, replacing the upstream traces.

Subsets that share a prompt version are independent samples of the same 49,133 rows drawn with that prompt at temperature 1 — Video_Games_reasoning, Video_Games_reasoning_sample2, Video_Games_reasoning_sample3; Video_Games_reasoning_v6, Video_Games_reasoning_v6_sample2, Video_Games_reasoning_v6_sample3, Video_Games_reasoning_v6_sample4, Video_Games_reasoning_v6_sample5. No reasoning text is repeated between them, so they can be used for majority voting, difficulty stratification, or augmentation. Rows align across every subset by _source_index.

Usage

from datasets import load_dataset

ds = load_dataset("mengdanzhu/genrec-dataset", "Video_Games_reasoning_v6_sample5", split="train")

Or point the training pipeline at this repo:

export SIDR_HF_REPO=mengdanzhu/genrec-dataset

Subsets

config	splits
Video_Games_reasoning	train
Video_Games_reasoning_sample2	train
Video_Games_reasoning_sample3	train
Video_Games_reasoning_v4	train
Video_Games_reasoning_v5	train
Video_Games_reasoning_v6	train
Video_Games_reasoning_v6_sample2	train
Video_Games_reasoning_v6_sample3	train
Video_Games_reasoning_v6_sample4	train
Video_Games_reasoning_v6_sample5	train
Video_Games_reasoning_v6_bestofn_K5	train

Video_Games_reasoning_v6_bestofn_K5 — best-of-N selected CoT

A filtered, single-trace-per-row subset derived from the five v6 samples (Video_Games_reasoning_v6, _sample2 … _sample5), not an independent draw.

Selection runs in two steps against a frozen Phase-2 v6 single-CoT scorer. For every candidate trace z_k the likelihood gap of the gold SID is scored as

delta_k = log p(y* | h, z_k) - log p(y* | h)

where the second term uses the empty <think></think> (non-thinking) block, so delta_k isolates the contribution of that one trace. Then per row:

1. best-of-N — keep the candidate with the largest delta_k (ties go to the earliest config);
2. reject — if even that best candidate has delta_k <= 0, the row is dropped entirely.

	rows
source rows	49133
accepted	47976
rejected (all 5 candidates delta <= 0)	1157 (2.35%)

Rows are spread across all five samples rather than concentrated on v6 (source 0: 10637, source 1: 9331, source 2: 9469, source 3: 9179, source 4: 9360), and the argmax buys 0.469 nats of mean delta over first-accept selection, changing the kept trace on 76.0% of rows.

Columns match the sibling v6 configs, plus provenance: _cot_source / _cot_config identify which sample won, and _cot_delta is its score. Rows align with every other subset by _source_index.

Generation prompts

The exact templates used to generate each subset live under prompts/. A prompt named prompt_v<N>.txt produced the *_reasoning_v<N> subset, so any subset can be regenerated or extended from the same instructions.

prompt	subset
prompts/prompt_v4.txt	Video_Games_reasoning_v4
prompts/prompt_v5.txt	Video_Games_reasoning_v5
prompts/prompt_v6.txt	Video_Games_reasoning_v6, Video_Games_reasoning_v6_sample2, Video_Games_reasoning_v6_sample3, Video_Games_reasoning_v6_sample4, Video_Games_reasoning_v6_sample5

All prompts emit strict JSON with predictability, transition_type and reasoning. v4 asks for a 4-8 sentence analyst monologue (evidence -> interest -> mechanism -> prediction); v5 tightens this to a 1-6 sentence evidence-grounded rationale and forbids reciting the full history; v6 returns to the 4-8 sentence analyst monologue (120-180 words) and additionally feeds in item descriptions, so each Semantic ID is characterised by its concrete attributes without ever naming the title.

Generation prompts

The exact templates used to generate each subset live under prompts/. A prompt named prompt_v<N>.txt produced the *_reasoning_v<N> subset, so any subset can be regenerated or extended from the same instructions.

prompt	subset
prompts/prompt_v4.txt	Video_Games_reasoning_v4
prompts/prompt_v5.txt	Video_Games_reasoning_v5
prompts/prompt_v6.txt	Video_Games_reasoning_v6, Video_Games_reasoning_v6_sample2, Video_Games_reasoning_v6_sample3, Video_Games_reasoning_v6_sample4, Video_Games_reasoning_v6_sample5

All prompts emit strict JSON with predictability, transition_type and reasoning. v4 asks for a 4-8 sentence analyst monologue (evidence -> interest -> mechanism -> prediction); v5 tightens this to a 1-6 sentence evidence-grounded rationale and forbids reciting the full history; v6 returns to the 4-8 sentence analyst monologue (120-180 words) and additionally feeds in item descriptions, so each Semantic ID is characterised by its concrete attributes without ever naming the title.

Checkpoints

Trained models live under checkpoints/<experiment>/<step>/ in plain HuggingFace format, so they load directly with AutoModelForCausalLM. The Stage-3 entries are GRPO checkpoints initialized from the matching Stage-2 CoT model, merged from their sharded FSDP form.

experiment	stage	checkpoint
Video_Games_stage2_cot_v4_gpt54_Qwen3-1.7B	Stage-2 CoT	final_checkpoint
Video_Games_stage2_cot_v5_gpt54_Qwen3-1.7B	Stage-2 CoT	final_checkpoint
Video_Games_stage2_cot_v6_single_Qwen3-1.7B	Stage-2 CoT	final_checkpoint
Video_Games_stage2_transition_cot_v6_Qwen3-1.7B	Stage-2 CoT	final_checkpoint
Video_Games_stage3_rl_no_kl_cot_v4_gpt54_Qwen3-1.7B	Stage-3 RL	global_step_200
Video_Games_stage3_rl_no_kl_cot_v5_gpt54_Qwen3-1.7B	Stage-3 RL	global_step_100

from huggingface_hub import snapshot_download
from transformers import AutoModelForCausalLM, AutoTokenizer

sub = "checkpoints/Video_Games_stage3_rl_no_kl_cot_v5_gpt54_Qwen3-1.7B/global_step_100"
path = snapshot_download("mengdanzhu/genrec-dataset", repo_type="dataset",
                         allow_patterns=[sub + "/*"]) + "/" + sub
model = AutoModelForCausalLM.from_pretrained(path)
tokenizer = AutoTokenizer.from_pretrained(path)

Checkpoints

Trained models live under checkpoints/<experiment>/<step>/ in plain HuggingFace format, so they load directly with AutoModelForCausalLM. The Stage-3 entries are GRPO checkpoints initialized from the matching Stage-2 CoT model, merged from their sharded FSDP form.

experiment	stage	checkpoint
Video_Games_stage2_cot_v4_gpt54_Qwen3-1.7B	Stage-2 CoT	final_checkpoint
Video_Games_stage2_cot_v5_gpt54_Qwen3-1.7B	Stage-2 CoT	final_checkpoint
Video_Games_stage2_cot_v6_single_Qwen3-1.7B	Stage-2 CoT	final_checkpoint
Video_Games_stage3_rl_no_kl_cot_v4_gpt54_Qwen3-1.7B	Stage-3 RL	global_step_200
Video_Games_stage3_rl_no_kl_cot_v5_gpt54_Qwen3-1.7B	Stage-3 RL	global_step_100

from huggingface_hub import snapshot_download
from transformers import AutoModelForCausalLM, AutoTokenizer

sub = "checkpoints/Video_Games_stage3_rl_no_kl_cot_v5_gpt54_Qwen3-1.7B/global_step_100"
path = snapshot_download("mengdanzhu/genrec-dataset", repo_type="dataset",
                         allow_patterns=[sub + "/*"]) + "/" + sub
model = AutoModelForCausalLM.from_pretrained(path)
tokenizer = AutoTokenizer.from_pretrained(path)