Upload folder using huggingface_hub

README.md

@@ -41,27 +41,27 @@ Several subsets under one dataset path. Pick the one that fits your use-case; re
 
 | Config | Rows | Schema | Best for |
 |---|---:|---|---|
-| `client_v1_full_corpus_raw` (default) | **23527** | full interaction (prompt + rawResponse + decision blob + call metadata) | failure-mode research, replay, end-to-end audit |
-| `client_v1_teacher_clean_raw` | **9937** | full interaction | fine-tuning, honest training-quality subset (single teacher model, current schema, non-stalled) |
-| `client_v1_teacher_clean_lean` | **9937** | derived per-decision (flat schema; see *Fields*) | quick analytics, lightweight loading, headline-statistics work |
-| `client_v1_26b_raw` | **1843** | full interaction | comparison: the `gemma-4-26b-a4b-it` MoE cohort alone (current schema, stalled decisions kept, 1 winning session) |
-| `client_v1_26b_lean` | **1843** | derived per-decision (flat schema) | comparison analytics: the MoE on matched game states |
+| `client_v1_full_corpus_raw` (default) | **28904** | full interaction (prompt + rawResponse + decision blob + call metadata) | failure-mode research, replay, end-to-end audit |
+| `client_v1_teacher_clean_raw` | **12021** | full interaction | fine-tuning, honest training-quality subset (single teacher model, current schema, non-stalled) |
+| `client_v1_teacher_clean_lean` | **12021** | derived per-decision (flat schema; see *Fields*) | quick analytics, lightweight loading, headline-statistics work |
+| `client_v1_26b_raw` | **1949** | full interaction | comparison: the `gemma-4-26b-a4b-it` MoE cohort alone (current schema, stalled decisions kept, 1 winning session) |
+| `client_v1_26b_lean` | **1949** | derived per-decision (flat schema) | comparison analytics: the MoE on matched game states |
 
 ```python
 from datasets import load_dataset
 
 # Default -- the full corpus, including failure modes
-full = load_dataset("chayuto/klondike-llm-decisions")  # 23527 rows
+full = load_dataset("chayuto/klondike-llm-decisions")  # 28904 rows
 
 # The training-friendly subset (filtered, single teacher)
-clean_raw  = load_dataset("chayuto/klondike-llm-decisions", "client_v1_teacher_clean_raw")   # 9937 rows
-clean_lean = load_dataset("chayuto/klondike-llm-decisions", "client_v1_teacher_clean_lean")  # 9937 rows, flat schema
+clean_raw  = load_dataset("chayuto/klondike-llm-decisions", "client_v1_teacher_clean_raw")   # 12021 rows
+clean_lean = load_dataset("chayuto/klondike-llm-decisions", "client_v1_teacher_clean_lean")  # 12021 rows, flat schema
 
 # The 26B MoE cohort on its own, for comparison
-moe_26b    = load_dataset("chayuto/klondike-llm-decisions", "client_v1_26b_raw")  # 1843 rows
+moe_26b    = load_dataset("chayuto/klondike-llm-decisions", "client_v1_26b_raw")  # 1949 rows
 ```
 
-The `client_v1_26b_*` configs are a **behavioural-comparison cohort**: all 1843 current-schema decisions from the `gemma-4-26b-a4b-it` MoE, with stalled/loop decisions deliberately kept (1115 of 1843 sit inside a stall, where the `teacher_clean` configs drop such rows). The corpus now holds 1 winning session for the 26B MoE; the cohort still centres on failure-mode contrast (stalled/loop decisions kept), so treat it as a behavioural contrast set more than as additional training data.
+The `client_v1_26b_*` configs are a **behavioural-comparison cohort**: all 1949 current-schema decisions from the `gemma-4-26b-a4b-it` MoE, with stalled/loop decisions deliberately kept (1221 of 1949 sit inside a stall, where the `teacher_clean` configs drop such rows). The corpus now holds 1 winning session for the 26B MoE; the cohort still centres on failure-mode contrast (stalled/loop decisions kept), so treat it as a behavioural contrast set more than as additional training data.
 
 ### Filtering by model
 
@@ -79,10 +79,10 @@ The `client_v1_teacher_clean_*` configs are already filtered to a single teacher
 
 Collected via an external client-side harness (closed-source) running the Klondike app and capturing every teacher-advisor call. Each game seeds a reproducible deal. Rows are deduplicated by their UUIDv7 `id` across re-exports; nothing is discarded.
 
-- **Collection window**: 2026-05-17 to 2026-06-19
-- **Sessions**: 134 distinct game sessions
-- **Models**: `gemma-4-31b-it` (19331), `gemma-4-26b-a4b-it` (1843), `gemini-3.1-flash-lite` (1661), `MiniMax-M3` (692)
-- **Schema tiers**: current (23156), legacy (371)
+- **Collection window**: 2026-05-17 to 2026-06-26
+- **Sessions**: 161 distinct game sessions
+- **Models**: `gemma-4-31b-it` (23235), `gemini-3.1-flash-lite` (2692), `gemma-4-26b-a4b-it` (1949), `MiniMax-M3` (1028)
+- **Schema tiers**: current (28533), legacy (371)
 
 ### Planned: `server_v1_*` configs
 
@@ -117,12 +117,12 @@ Derived per-decision rows, flattened. Built by joining each successful interacti
 
 | Move type | Count | Share |
 |---|---:|---:|
-| `draw_card` | 11673 | 50% |
-| `tableau_to_tableau` | 6685 | 28% |
-| `tableau_to_foundation` | 2060 | 9% |
-| `discard_to_tableau` | 1225 | 5% |
-| `discard_to_foundation` | 949 | 4% |
-| `recycle_stock` | 935 | 4% |
+| `draw_card` | 13545 | 47% |
+| `tableau_to_tableau` | 8874 | 31% |
+| `tableau_to_foundation` | 2642 | 9% |
+| `discard_to_tableau` | 1553 | 5% |
+| `discard_to_foundation` | 1159 | 4% |
+| `recycle_stock` | 1131 | 4% |
 
 ## Failure modes are a feature of `*_full_corpus_raw`, not a bug
 

client_v1_26b_lean/part-00000.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a5ad6842f64db66c48431c6c31560e551d5a146bdd7e34b1d155f9e14be393fa
-size 9380085
+oid sha256:19dc9e68ad63a86131a83dec021936b63fff2eac6dc488d624a38d26a067ce52
+size 9956665

client_v1_26b_raw/part-00000.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:113e8a762256dea7f9c315d25f4f37eba3bbf588ba36c67f0e85d858c2651f44
-size 10030374
+oid sha256:829949eb07b02d6698ccae2a2a8a74c67d9d1d92cad9cd8440001511423c3a88
+size 10638451

client_v1_full_corpus_raw/part-00011.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c4e39982e934d10595f4c6377a368b6bd47102555f60fb48627bd790e550527a
-size 6781218
+oid sha256:b27d10e8f191a23d84afcbfa80da3704b273efe1ea105e470861427fddbc860a
+size 9372094

client_v1_full_corpus_raw/part-00012.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:73975af00dc72962c4e7ef271878262024e998901c311b7c4347de61594b8a8e
+size 6758397

client_v1_full_corpus_raw/part-00013.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6f5583eaeaa15bf8c7b1007dd0b2170f66cbf2aaa6613cb7066bace3348db053
+size 9161333

client_v1_full_corpus_raw/part-00014.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1db1d42c598f3a5b0a24f22c534076d15676bb9ce7c9da88a93630babd48a3f5
+size 3407982

client_v1_teacher_clean_lean/part-00004.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:db1e2acc92dca916a3a298a491a0c05956078ae1986e576b22eeab34940b08ba
-size 6109416
+oid sha256:4b49b885f006ad1b8c49adbe2ba2d60f09571f722dac09fac2af153a71eccaa3
+size 6264635

client_v1_teacher_clean_lean/part-00005.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:83e13fbb8f7b5302b97f91f21989f3243803e420b152fc143e7d10d6a263b222
+size 6223576

client_v1_teacher_clean_lean/part-00006.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:61a46d55f486642561a2a843df5553646792f1b8016dec5533025a5f6a12be94
+size 66676

client_v1_teacher_clean_raw/part-00004.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:83b9bc5de618db02524b13c284bf91e2e25bb2ba2e1ca5b6e4cebdc2656bafbd
-size 6932689
+oid sha256:98f206fdaaed15e3caef1add74ff01c97249d2fb2d20e0a5bc325e7fb387c885
+size 7108030

client_v1_teacher_clean_raw/part-00005.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:285aa95c9f0ca6e2c9010ad6a86301d14c80cc31a5ac9c1289399d5e28202412
+size 7070338

client_v1_teacher_clean_raw/part-00006.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:221dbe7f247eb5afbaaee8a11e2be53bd892601ee5bd6daeca2906ea7af15dae
+size 88641