mzio/aprm-sft-thoughts-snorkel-insurance-base_best-adamw30-lp0

Act-PRM SFT thoughts — snorkel-insurance insurance

Act-PRM (Action Process Reward Models) infers the latent thoughts behind logged, action-only agent demonstrations via an offline EM. For each logged action x in state s we sample G=4 candidate thoughts z, score each by the length-penalized action likelihood

reward(z) = p(x | s, z)

(len_frac grows with the thought's token length), and mark the best thought (argmax reward). The (thought + action) span is then what downstream SFT / RL trains on.

This dataset publishes the full E-step: every one of the G sampled thoughts per logged action, each with its reward, likelihood (p(x|s,z)) and thought_tokens, plus the best_index. That is, it is not reduced to the top-1 — you can re-derive top-1, top-half, EM-weighted, or any other selection downstream (see below).

Variants (files / splits)

file / split	scorer	em_checkpoint	relabel run
base_best.jsonl	base	best	insurance_s1relabel_base_adamw30

• scorer — which model computed p(x|s,z) in the E-step: the trained policy LoRA, or the frozen base model (--score_with_base).
• em_checkpoint — which EM training checkpoint the scorer was resumed from: the best-metric step, or the last step.

Every row also carries scorer / em_checkpoint columns, so you can also concatenate the files and filter.

Note: train-split rows have the full G=4 sampled thoughts; the held-out eval-split rows were relabeled with a single sample (G=1), so their thoughts/rewards lists have length 1 and best_index == 0.

Schema (one row per relabel-variant x logged action-step)

field	type	meaning
domain	str	{domain}
scorer	str	policy or base
em_checkpoint	str	best or last
split	str	train or eval (eval = held-out; G=1)
uid	str/int	source trajectory id
sample_id	int	dataloader sample id within the relabel pass
timestep	int	index of this action among the trajectory's actions
system_prompt	str	the agent system prompt
messages	list[{role, content}]	state s: context up to (not incl.) the action
target_action	str	the logged ground-truth action x (a <tool_call>...</tool_call>)
thoughts	list[str]	the G sampled candidate thoughts z
likelihoods	list[float]	`p(x
rewards	list[float]	length-penalized reward for each thought
thought_tokens	list[int]	token length of each thought
best_index	int	argmax-reward thought index

Selecting thoughts

from datasets import load_dataset
ds = load_dataset("mzio/aprm-sft-thoughts-snorkel-insurance-base_best-adamw30-lp0", split="policy_best")   # or base_best / policy_last / base_last
row = ds[0]

# top-1 (what the paper's SFT commits): the best length-penalized thought
top1 = row["thoughts"][row["best_index"]]
assert row["rewards"][row["best_index"]] == max(row["rewards"])

# the SFT target span is thought + action:
sft_target = top1 + "\n\n" + row["target_action"]

# top-half: keep the thoughts whose reward is in the top 50%
import numpy as np
order = np.argsort(row["rewards"])[::-1]
top_half = [row["thoughts"][i] for i in order[: max(1, len(order) // 2)]]

# EM weights: group-normalized (softmax-like) weights over all G thoughts,
# e.g. a temperature-1 softmax over rewards (or normalize exp(likelihood)):
import math
r = row["rewards"]
Z = sum(math.exp(x) for x in r)
em_weights = [math.exp(x) / Z for x in r]   # weight every (thought+action) span

Provenance

Generated by Act-PRM Stage-1 relabel passes over snorkel-insurance insurance expert demonstrations (no_train generate-only passes), model hf_qwen3_4b_instruct, group_size=4, length_penalty=0.0. The held-out RL-eval tasks are excluded. Built with scripts/export_sft_dataset_hf.py.

The LoRA that scored p(x | s, z) was itself trained by the Stage-1 EM:

EM setting	value
optimizer	adamw
learning_rate	4e-05
num_batches	30
advantage_mode	action_probs
length_penalty	0.0
lora_config	r32_a32_linear