kimperyang commited on
Commit
4277504
·
verified ·
1 Parent(s): 7a90abf

K-39 RFT records verified_000: 15 self-rollout records (E2 / plan A, training worlds)

Browse files
rft/selfrec-r1/README.md ADDED
@@ -0,0 +1,29 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # rft/selfrec-r1: verified self-rollout records of the 4B students (K-39)
2
+
3
+ Training records for rejection-sampling fine-tuning (RFT): the successful episodes of the students' own rollouts in our
4
+ training worlds (`../../rollouts/e2-train-r1/`, `../../rollouts/e2negft-train-r1/`), H100 cluster, 2026-10-27
5
+ (docs/sync/katrina1.md K-39, K-40; repo KimperYang/VLMAuditor).
6
+
7
+ ## Selection
8
+
9
+ `tools/worldgen/mcpgen_validate.py --no-judge` over every farm run, the acceptance of the 10-25 fourth batch. An episode becomes a
10
+ record only if it passes all checks:
11
+ - score: exactly one final report; its category is the task's; its cited frames show the bug (oracle: target pixels / oracle
12
+ frames / before-after for change bugs); its description names the target object;
13
+ - the model ended the episode itself with an accepted `done`; actions within the 40-action budget; no lost WebGL context and no
14
+ flat main frame; no hidden-guidance leak; every tool call valid against its schema; every image on disk.
15
+ The text-only report judge of earlier batches was not run (its server is down); it used to drop about 5% more.
16
+
17
+ ## Files
18
+
19
+ - `verified_NNN.jsonl`: the records that are new in that upload, format `mcp-v1` (`tools` + `messages`, as E2's records).
20
+ - `source` is `selfrec`; `self_model` is `e2` (E2 ckpt-394) or `e2negft` (plan A ckpt-263); `run`, `task`, `variant`,
21
+ `world`, `family`, `code` identify the episode. Each task has at most two episodes (two spawns), so at most two records.
22
+ - `llm.base_url` removed (local server address).
23
+ - `images_NNN.tar`: every image those records reference, under the same relative paths (extract into one IMAGE_ROOT and train
24
+ with it: paths look like `out/mcpeval/<run>/solo/<episode>/mcp/frames/a000.jpg`, plus the in-context example images
25
+ `examples/icl/images/...`).
26
+ - `verify_stats_NNN.json`: the validator's counts over all episodes at that time (outcome per first failed check, per model).
27
+
28
+ The episodes were played in the evaluation format (no hidden block, 640 px, temperature 0.7, top_p 0.8, top_k 20, thinking off,
29
+ 40 actions). Our own procedurally built worlds only; no benchmark tasks, rubrics or answers.
rft/selfrec-r1/images_000.tar ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1850096d2a65c7a814848081a2af2c0bba729157613c16db77c25dca838237f0
3
+ size 66048000
rft/selfrec-r1/verified_000.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
rft/selfrec-r1/verify_stats_000.json ADDED
@@ -0,0 +1,54 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "episodes": 180,
3
+ "outcome": {
4
+ "score": 165,
5
+ "verified": 15,
6
+ "turns_masked_invalid_call": 1
7
+ },
8
+ "by_model": {
9
+ "k38-n0": {
10
+ "score": 165,
11
+ "verified": 15
12
+ }
13
+ },
14
+ "verified_composition": {
15
+ "code": {
16
+ "G1": 1,
17
+ "G2": 3,
18
+ "G3": 1,
19
+ "S2": 1,
20
+ "T1": 1,
21
+ "T3": 5,
22
+ "V3": 1,
23
+ "V4": 2
24
+ },
25
+ "model": {
26
+ "k38-n0": 15
27
+ },
28
+ "spawn": {
29
+ "mid": 10,
30
+ "random": 5
31
+ },
32
+ "reasoning": {
33
+ "null": 15
34
+ },
35
+ "actions": {
36
+ "11-20": 2,
37
+ "21-30": 1,
38
+ "31-40": 12
39
+ },
40
+ "source": {
41
+ "teacher": 15
42
+ },
43
+ "source_code": {
44
+ "teacher:G1": 1,
45
+ "teacher:G2": 3,
46
+ "teacher:G3": 1,
47
+ "teacher:S2": 1,
48
+ "teacher:T1": 1,
49
+ "teacher:T3": 5,
50
+ "teacher:V3": 1,
51
+ "teacher:V4": 2
52
+ }
53
+ }
54
+ }