File size: 1,654 Bytes
5290c1b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
# missing

Teacher rollouts (k=3, temperature 0.8, max_tokens 1792) with thought echo scores for
49,341 of the 49,341 Affine corpus-D turn_ids in
[`vuhaian/49k_missing`](https://huggingface.co/datasets/vuhaian/49k_missing) (`turn_ids.txt`),
at corpus epoch 40 (manifest `fe61418440d6`, view `duel_turns@v4`).

## Generation

- Teacher `Qwen/Qwen3.8-27B@1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0`, vLLM 0.28.0, TP=1 x 32 replicas (4 x 8 H200)
- Prompt `evalsrv.chat.gen_prompt` (ends inside `<think>`); 3 samples per turn in one `n=3` request,
  max_tokens = `[duel]` max_thought_tokens 1024 + max_action_tokens 768 = 1792
- Split `evalsrv.chat.split_rollout`: z = thought, y = action
- Echo: `lp_thought` = lpC(z|x), mean logprob per byte (`VllmModel.score_thought`);
  `lp_own` / `lp_empty` are not computed (null)

Per turn: `turn_id, action_kind, source, stratum, corpus_epoch, prompt_tail, max_tokens,
temperature, n_parsed, rollouts[3]`. Per rollout: `raw` (completion as sampled), `reasoning`,
`content`, `z`, `y`, `parsed`, `finish_reason`, `prompt_tokens`, `completion_tokens`,
`think_closed`, `lp_thought`, `lp_own`, `lp_empty`. Prompts are not stored.
Rows follow the order of `turn_ids.txt`.

## Files

- `49k_missing_e40.jsonl`: 49,341 turns
- `49k_missing_e40.meta.json`: teacher / sampling / echo / corpus stamps
- `49k_missing_e40.errors.jsonl`: 0 turn_ids without a row ({})

## Stats

- 148,023 rollouts: 92.1% parsed, 89.7% closed `</think>`,
  12.5% hit max_tokens; 2,043 turns with 0/3 parsed
- mean prompt 4,337 tokens, mean completion 513 tokens
- action_kind: bash 24,551, tool_call 15,097, text 8,674, terminus_json 880, boxed 139