ReadAll / ReadTwice: historical HQA step15
This public repository contains the historical HQA-stage step15 checkpoint
from readall_sft_hqa_rl_20260917_r1, saved on September 18, 2026.
The checkpoint has 120 optimizer updates and 960 effective question groups.
The September 21 mixed DocQA+HQA v2 stage is a separate training run.
All 339 BF16 tensors were checked against all eight original FSDP model shards
before upload. Four safetensors weight files, the index, tokenizer, and original
model and generation configurations are included. See checkpoint_manifest.json
for the weight verification and file hashes.
Historical greedy HQA results
| Context length | SubEM | EM | F1 |
|---|---|---|---|
| 56k | 78.91% | 62.50% | 75.27% |
| 112k | 82.03% | 64.06% | 78.96% |
| 224k | 75.00% | 60.16% | 73.20% |
Each length contains all 128 fixed development examples. Saved predictions were
re-scored before publication. See evaluation.json for exact evaluation settings.
No LongBench score is claimed for this checkpoint.
These results use greedy decoding (temperature=0, top_p=1, seed=801), paged ReadTwice, 5,000-token chunks, balanced page-local outlines with at most 24 sections per page, and serial memory across pages. Historical memory handling uses the strict complete-rewrite protocol, SKIM=512, UPDATE generation cap=1088, FINAL=1024, and a 1024-token memory limit. The later MemAgent-compatible UPDATE=1024 protocol differs; cross-protocol scores are not a controlled checkpoint comparison.
The native model context is 32,768 tokens. Long inputs are processed through the
ReadTwice workflow with paging; they are not passed to a single model call.
The original generation configuration retains sampling defaults. Override with
do_sample=False for greedy inference.
from transformers import AutoModelForCausalLM, AutoTokenizer
repo_id = "Xirui1208/readall-readtwice-hqa-step15-20260918"
tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForCausalLM.from_pretrained(
repo_id, torch_dtype="bfloat16", device_map="auto"
)
The repository is public and can be downloaded without authentication.