Upload HISTORY.md with huggingface_hub

HISTORY.md

@@ -8,7 +8,7 @@ Last updated: 2026-08-08.
 
 ## One-line summary
 
-We built a **~1.7B Llama-adjacent PCS decoder**, FineWeb ~**8B** tokens, chat-SFT through **ground-adhere** (ship Gradio **6/9/9**). **Product north star (clarified 2026-08-08):** harness agent — tools → call → answer **from tool results**; not parametric knowledge. Chron/facts-synth failed (wrong teachers). Attention-mass (§31): not attention-blind. **Hermes-tools smoke** in flight (§33) after fixing Hub-stream GPU starvation (async per-stream refill + PrefetchBatcher → ~8–9k tok/s, GPU pegged). Ship stays until tool-loop gate. Scripts under **`scugnizz-llama-pcs/training/`**.
+We built a **~1.7B Llama-adjacent PCS decoder**, FineWeb ~**8B** tokens, chat-SFT through **ground-adhere** (ship Gradio **6/9/9**). **Product north star:** harness agent — tools → call → answer **from tool results**. Hermes-tools **smoke** done (§33; `tool_context_eval` **2/6**, first tool traces). **Full hermes-tools meal** next (§34). Ship stays until tool-loop gate. Scripts under **`scugnizz-llama-pcs/training/`**.
 
 ### Current ship (do not overwrite lightly)
 
@@ -630,6 +630,23 @@ Passage grounding / “grounded synthesis” remain useful **proxies** for “re
 
 **Do not ship** until smoke finishes healthy + tool-loop eval; ship remains ground-adhere.
 
+### 34. Hermes-tools full meal (2026-08-09 — **in flight**)
+
+Smoke (§33) finished: train healthy; `tool_context_eval` **2/6** (ship/tool-context was **4/6**). Failures often showed **`<tool_call>` traces** — first time this weights path saw tools; call CE stuck, answer-from-tool underfed (only **10%** tool_context).
+
+**Meal recipe** (profile `chat-hermes-tools` / `examples/train-runpod-hermes-tools.sh`)
+
+| Item | Value |
+|------|--------|
+| Init | **ship** ground-adhere (not smoke — avoid locking call-spam) |
+| Mix | same 4 Hermes repos + `__tool_context_mix__` at **0.35** (was 0.10); hermes weights 0.20/0.15/0.12/0.18 |
+| Steps / LR | **3500**, **5e-6→1e-6**, warmup 80, pulses **1000/2000/3000** |
+| Batch | **10×accum 2** (same tok/step as smoke 5×4) |
+| Filters | `REJECT_TOOL_TRACES=0`, `ALLOW_PHANTOM_TOOLS=1`, no phantom-thinking reject |
+| Live | pod **`muu8k7sau5x3jb`** (~$1.69/h), Hub `training-runs/sft-chat-v2-hermes-tools-8b-20260809-052231` |
+
+Gate: `tool_context_eval` ≥4/6 (+ harness later). Ship stays ground-adhere until then.
+
 ### 23. Why answerable-calibrate (2026-08-06 — **finished**)
 
 **Problem after wiki-balance:** refusal **9/10** (keep), wiki adherence **4/10** (mostly abstain-on-answerable), length **8/9**, `tool_context_eval` **2/6**. Tool-call CE deferred.
@@ -733,7 +750,8 @@ Published curated train/eval/Gradio package to `ProjectScugnizz/scugnizz-llama-t
 | adh-swing | `…/adh-swing-8b-20260807-121936` | no ship (Gradio ≤ ship; Jobs peaked mid then fell) |
 | chronicling | `…/chronicling-8b-20260807-205348` | no ship (Jobs 6/0/2; Gradio 0/10/1) |
 | facts-synth | `…/facts-synth-8b-20260808-064616` | no ship (Gradio 5/9/8; Jobs 5/0/7) |
-| hermes-tools-smoke | `…/hermes-tools-smoke-8b-20260808-183916` | **in flight** (§33); tool-call CE on; not ship yet |
+| hermes-tools-smoke | `…/hermes-tools-smoke-8b-20260808-192108` | snack done; `tool_context_eval` **2/6**; first tool traces |
+| hermes-tools | `…/hermes-tools-8b-20260809-052231` | **in flight** (§34) full meal |
 
 ---
 
@@ -779,16 +797,17 @@ Published curated train/eval/Gradio package to `ProjectScugnizz/scugnizz-llama-t
 38. **Hermes-tools smoke first** — ship init, `REJECT_TOOL_TRACES=0`, Hermes Hub traces + tool_context snack, ~800 steps; prefer RunPod over HF Jobs for cost (§33).
 39. **SFT Hub streams must not sit on the train critical path** — per-stream async refill + PrefetchBatcher; CPU “idle” was network-wait. Phantom-tool / phantom-thinking filters can starve tool-call smokes — loosen for smoke when CE-on is the point.
 40. **Tok/s next: microbatch↑ / accum↓ (and maybe no grad-ckpt), not more prefetch** — once GPU ~95%+, data prebuffer is done (§33).
+41. **Smoke ≠ meal** — 800-step hermes smoke proved call CE; `tool_context_eval` 2/6 is expected first-contact. Full meal: more `__tool_context_mix__`, longer steps, ship init (§34).
 
 ---
 
 ## Open next steps
 
-- **Ship remains** ground-adhere (`…/sft-chat-v2-ground-adhere-8b-20260807-053640`) until a **tool-loop** gate beats it.
-- **Finish hermes-tools smoke** (`…/hermes-tools-smoke-8b-20260808-183916` / pod `9os3s41pwo312z`); confirm healthy loss + no OOM; then **full Hermes continue** (~3–4.5k steps) with same data path.
-- Bake **BATCH=10 GRAD_ACCUM=2** (same effective tok/step) ± no-grad-ckpt probe into next launch (§33 squeeze).
+- **Ship remains** ground-adhere until a **tool-loop** gate beats it.
+- **Finish hermes-tools meal** (`…/hermes-tools-8b-20260809-052231` / pod `muu8k7sau5x3jb`); gate `tool_context_eval` ≥4/6.
 - Gate on harness / `tool_context_eval` / agent traces — **not** wiki adh.
 - **no-context / question-only → abstain** snack (#27) still useful (no tool yet / empty result).
 - More FineWeb / HQ anneal only if the **agent loop** fails for lack of language/plasticity, not for lack of trivia.
 - LongRoPE / longer ctx only if agent traces outgrow 4k.
+- Optional: MiniMind-style Agentic RL after SFT tool loop is healthy (parked reference).
 - Optional: finer attn probe only if architecture suspicion returns.