README.md

# **Qwen3.5-35B-A3B — v0.9 dataset**

Last substantive update: 2026-08-04

**v0.9 is v0.7 with 38,145 assistant turns excluded from the loss. Nothing else
differs — same rows, same order, same content, same reasoning, same pack seed.**

The design and its evidence are in `QWEN35_V0_9_DATASET_PLAN.md`. This file records
what was actually built, where it is, and how to check it.

---

## **Location — b200 cluster**

    /NHNHOME/WORKSPACE/wbl-workspace/agentic-rl-260803/datasets/v0.9/
      raw/v0_9_final.jsonl              742,665 rows, 34,528,993,378 bytes
      packed-65k/train/                 137,384 packs in 8 parquet shards
      provenance/mask-flags-v1.jsonl    38,145 records, one per masked turn
      provenance/v0_9_build_report.json
      provenance/v0_9_verify_report.json
      provenance/census_v0_9_v2.json
      provenance/*.py, *.sbatch         the exact scripts that produced it
      logs/

| artifact | SHA-256 |
|---|---|
| base `v0.7/raw/v0_7_final.jsonl` | `cd94b138f7988b81c48e12ff3b75559192ef7c4b48fdc2f86d076e48d830266d` |
| **`raw/v0_9_final.jsonl`** | `7f742b2b3c3ab569c007070ca7d17089cc03d1cf4fa9375cccd58dc305b62a2f` |
| staged v0.8 classifier flags | `a5d2f1f5999e7f03b495ff3e389656574c89a4e2830cb1e7bac3d0777299a0c1` |

Superseded artifacts from the abandoned `--claim-policy correct` build are kept under
`superseded-correct-policy/`; they are not inputs to anything.

---

## **What is masked**

38,145 turns across 33,291 rows — **5.1% of assistant turns in 4.5% of rows**. Reasons
overlap (most stalls are also malformed JSON), so each masked turn carries the set of
reasons it qualifies for. The per-reason totals sum to 66,420; the union is 38,145.

- **32,836 `terminus-json`** — the Terminus-2 harness rejected the action object
- **28,292 `stall`** — empty content (28,270), no `commands` key (20), empty mini-swe
  content (2)
- **3,116 `false-claim`** — final assistant turn asserting `task_complete: true` on a row
  a trusted verifier failed (numeric non-bool reward, explicit `passed`, no infra
  exception, not a known-broken verifier family). 84 further candidates also carried an
  executable command and were left alone
- **1,385 `toolcall-syntax`** — malformed or schema-invalid tool call; 484 land on the
  general-agent sources τ³ measures (API-Bank 478, Toucan 5, ToolACE 1)
- **791 `prose-command`** — mini-swe turns that write a command as prose instead of
  executing it (727 of 738 `no-toolcall`, 64 of 305 `trailing-after-submit`)

Combination counts: `stall+terminus-json` 28,275 · `terminus-json` alone 4,561 ·
`false-claim` 3,116 · `toolcall-syntax` 1,385 · `prose-command` 791 · `stall` alone 17.

## **What is deliberately left trained**

- **9,692 `rewrites`** — v0.7 has no masks and this is the most error-reactive class
  (58.66%), while τ³ error recovery is the known weakness
- **252 prose-only mini-swe turns** — end-of-run summaries with no command text; v0.7
  trained them normally
- **200 `terminus-unparsable-harness-rejected`** turns that carry no `terminus-json`
  class. The other 4,264 unparsable turns do carry it and are masked
- **796 rows that passed** contain 819 masked turns. This is intended: a stall inside a
  trajectory that eventually passed is still a stall

---

## **How it was built**

    workflows/sft/v0_9/census_v0_9.py     measure only, never mutates
    workflows/sft/v0_9/build_v0_9.py      one streaming pass over the base
    workflows/sft/v0_9/verify_v0_9.py     independent on-disk re-check
    workflows/sft/v0_9/build_v0_9.sbatch  b200 launcher (1:56 wall)
    workflows/sft/v0_9/pack_v0_9.sbatch   b200 pack at 65,536

Packed with the same artifacts that produced v0.7 and v0.8 on b300 — packer
`976ce4bb…`, `tokenizer.json` `06b95093…`, `chat_template.jinja` `f8a27a12…`, staged at
`agentic-rl-260803/env/qwen35-qwen36-thinking-generation-spans`. A 300-row sample
tokenizes to an identical digest on both clusters
(`1594220b35a637c8…`; transformers 5.8.1, tokenizers 0.22.2, megatron-bridge 0.5.0 on
each), so the b200 pack is not a different rendering of the corpus.

**Seed 20260731 and 8 shards, matching v0.7**, on v0.7's row order — shard membership is
identical, so a v0.7↔v0.9 comparison is not confounded by a different packing.

## **Pack result, against v0.7**

| | v0.7 | **v0.9** | delta |
|---|---:|---:|---:|
| rows in | 742,665 | 742,665 | 0 |
| rows packed | 742,661 | **742,656** | **−5** |
| packs | 137,387 | **137,384** | −3 |
| tokens | 8,978,973,597 | **8,978,945,130** | −28,467 |
| packing efficiency | 99.71% | **99.73%** | — |
| rows over 65,536 tokens | 4 | **4** | 0 |
| one epoch @ GBS16 | — | **8,587 steps** | — |

**The whole difference is 5 rows, and they are accounted for.** `rows_over_max_len` is 4
in both, which independently confirms no content grew or shrank; the extra 5 dropped rows
are `rows_fully_masked` — trajectories in which *every* assistant turn was a defect, so
nothing trainable survived. The packer drops those rather than emit a zero-loss sample.

Loss coverage: **54,577,131 masked tokens against 3,408,327,232 trainable** — masking
removes **1.58%** of the assistant-target loss. Span integrity was clean:
`anomalies=0`, `turn_flags_on_non_assistant=0`, `span_turn_count_mismatch=0`,
`span_bad_start=0`, `span_multiple_headers=0`, so every mask landed on the token span it
was meant to. The validator additionally sampled 128 packs and read back 3,001,541
assistant loss tokens.

## **Gates that passed**

Build (`v0_9_build_report.json`):

- base and flag digests re-checked at build time; output digest recorded
- 742,665 rows in, 742,665 out; no rows added, dropped or reordered
- every modified row compared against source-plus-declared-changes — an unintended edit
  anywhere is a nonzero exit; 0 occurred
- 0 alignment anomalies: every flag resolved to an assistant turn
- per-reason counts equal to the census figures, exactly
- 4,963/4,963 untouched-row probes re-serialise byte-identically

Verify (`v0_9_verify_report.json`), separate logic reading both files off disk:

- **0 content corrections seen** — the only field that changed anywhere is
  `step_loss_mask`
- the base carried no `step_loss_mask` of its own, so all 38,145 are this build's
- every masked turn's source content still matches the fingerprint recorded at build time
- every `false-claim` mask is the final assistant turn of a `passed == False` row whose
  parsed object carried `task_complete: true`
- 0 errors

Pack (`packed-65k/validation-summary.json`), `validate_qwen35_packed_dataset.py`:

- 742,656 of 742,665 samples present, 9 dropped and both causes accounted for
- 137,384 packs at 65,536 tokens, 99.7262% efficiency
- 0 span anomalies of any kind; masks land on the intended token spans

## **Training**

3 epochs, TP2/EP8, MBS1/GBS16, LR 1e-5/1e-6, save every 500 steps. One epoch is **8,587
steps** (137,384 packs / GBS 16) — the same as v0.7, whose 137,387 packs also round to
8,587 — so matched-step comparisons against v0.7 are exact. Two epochs is **iter 17,174**
and three is **25,761**; with saves every 500, the nearest checkpoints are 17,000 and
25,500. Point `packed_data_path` at `datasets/v0.9/packed-65k`. Evaluation contract,
hard floors and the checkpoint-selection rule are predeclared in
`QWEN35_V0_9_DATASET_PLAN.md` §5 — read it before scoring, not after.