Hkang/litecoder-terminal-sft-qwen3-8b-embeddings

LiteCoder Terminal SFT — Qwen3-8B state/action embeddings

Complete precomputation of 308,641 assistant steps from 11,255 trajectories in Lite-Coder/LiteCoder-Terminal-SFT. Source revision: 6acdbbdb29979e4b8ea717b12accc8214606d087. Source SHA-256: 50f58e33b4caaac73e6e12afe1a90206d18e4f47080da8b0b7d9a704a0c7c43f. Encoder: Qwen/Qwen3-8B at b968826d9c46dd6066d109eabc6255188de91218.

Encoding contract

Each state contains only messages BEFORE its assistant action, rendered with the encoder's chat template (add_generation_prompt=False, default thinking template behavior). Each action is the complete raw assistant response, including any reasoning/JSON/tool text, tokenized without special tokens. States retain the last 8,191 tokens; actions retain the first 8,191. Truncation counts are in metadata. Last-token pooling of the final hidden layer, L2 normalization, 4,096 dimensions. Inference uses BF16; stored vectors use FP16. No projection head is applied.

Files and use

train/ and val/ each contain state_embeddings.pt, action_embeddings.pt and metadata.json, directly compatible with jackyk02/contrastive_learning/train/contrastive_train_embeddings.py. Rows in both tensors correspond exactly to metadata.json["samples"].

import json, torch
states = torch.load("train/state_embeddings.pt", weights_only=True)
actions = torch.load("train/action_embeddings.pt", weights_only=True)
samples = json.load(open("train/metadata.json"))["samples"]

Split assignment is deterministic: SHA-256 of the first user message, first eight hex digits modulo 10; bucket 0 goes to validation. Repeated instructions stay in the same split. This is a text-identity split, not a claim of semantic task deduplication. Train: 278,817 steps. Validation: 29,824 steps.

The source has no outcome scores: reward and generator model stay null. These are curated SFT examples, not independently verified successes. No hard negatives or new agent rollouts were generated.

manifest.json pins source/model revisions; validation.json records exact coverage, checkpoint checksums, finite/unit-norm checks and truncation counts. pipeline/ contains the scripts used to build this artifact.